I agree the non-fungible nature of a H100 (or whatever) is a massive challenge. How do you actually verify usage/thermal history on these? Hours run and thermal violations aren't stored on the card IIRC, only ECC error counts and retired pages persist in the GPU's own InfoROM, host side monitoring like DCGM logs, are only as good as whatever the seller hands over. Is condition at inspection based on a real monitoring export from the deployment or just an InfoROM/diagnostic check at the time, and if it's the latter isn't hours run and thermal history basically unverifiable unless we assume good faith and the seller was logging the whole time ?
Essentially how do you avoid having the 'used car problem' without leaning heavily on seller reputation & warranties?
Great question. We don’t treat “used” as one bucket. Before the RFQ goes out, the buyer says what logs and tests they need. If the seller can’t provide usage history, that’s stated in the quote rather than inferred. The market then prices that difference. Hardware with better records and more thorough testing should attract higher bids.
I assume this is a strictly B2B proposition in spite of "built for every side of the GPU market" -- meaning I'm probably not going to be able to buy/bid on small (e.g. 2-8) lots of used, matched RTX 3090s for example?
For now, we are focusing on institutions, since most of this hardware has strict export controls. We deal with all sizes though, we can do a single card to 1024 servers. Can always reach out to us to see if a specific request would fit in the marketplace or not
I’m very curious about how you’re fending off fraud— just wondering about how a lot of dark web transactions supposedly have escrow, but people waive it and then get burned. Do you have anything beyond KYC/escrow?
We use Stripe Connect to transfer the funds, and it is not a waivable process. In addition to KYC documentation, the team personally meets with any participant we onboard to the marketplace. We also have the ability to manually intervene in transactions if need be, but so far we have not had any transaction issues. It almost reminds me of how people were super skeptical of Uber and Airbnb when they first launched, but human trust goes a long way, sometimes further than we originally anticipate.
One thing that guides how we think about this is: we will never ask you to make the payment before you feel comfortable with the counterparty you are dealing with.
But what about the GPUs that are considered defective for corporate and enterprise use, but perfect for personal use...say like A100s that maybe having ECC issues but I can live with that
We ask any seller to post the condition and testing reports for the hardware they sell, so there is no info asymmetry. As long as there is a willing buyer and a willing seller at the conditions specified, transaction goes through! We deal with A100s all the time.
Nice thing about these chips is that they have built in software that allows you to see the condition. The secondary market also supports the circular economy, most of these servers are proving to be useful longer than what the market initially thought. There is always a case to go with used equipment instead of new ones.
As an individual, will I have access to pricing or is that limited to institutions as well?
For that matter, how transparent is this marketplace? Do I need to put in an RFQ for a specific SKU to get a point-in-time price or do I see the whole marketplace?
You mention $300 million dollars in RFQs, but how much of that resulted in chips and cash trading hands? What stops someone from offering the seller a better price to continue the transaction outside of Stoa?
Great questions. We do not accept individual accounts today. Marketplace access and pricing are limited to institutions.
A buyer can create an RFQ for bare GPUs or servers, specifying required configurations while leaving others open. They can see an indicative price range that updates with market activity, then compare all supplier quotes. Other participants’ activity stays private.
Our RFQ activity came from institutions we worked with directly. GPU transactions can take weeks due to sourcing, KYB, financing, and logistics. We are now onboarding those institutions so more of this activity can move through our marketplace.
We create value by making transactions easier for both parties. We enable buyers to compare quotes in one place, while suppliers receive qualified demand and market intelligence without exposing their bids. Binding quotes and non-circumvention terms also help this process.
I anticipate some kind of futures on GPUs based on a number of factors such as the availability for rent, the availability at your marketplace, and the models currently available for self-hosting.
Yes, those types of differences across different providers is what the finance world calls basis risk. As more and more transactions take place, we hope to bring transparency into how different components/specs can change pricing, and make them tradable too.
Basis risk is also a big part of futures markets, how much does spot price differ from forward pricing and why. For physical GPUs, theory would expect a downward sloping forward curve due to the depreciation factor, but from what we have seen recently, sometimes not the case!
Those work with rental agreements on compute, we directly transact the hardware itself. Instead of per hr pricing, we do physical unit prices. We help with the shipment and clearing process too.
Yep, we currently only transact within the US in the marketplace. We handle end user disclosure type checks on the platform. For any international deals, we still support them via our over-the-counter brokerage arm today.
Is pricing spot-only or do you support forward commitments? GPU supply deals seem to want term structure - a lab locking capacity for 6 months cares about a very different price than burst spot. Curious how you handle default risk on the longer terms.
We do both! Solving for the forward curve has very interesting applications in the GPU market, as depreciation is a big factor.
We have credit limits/margins in place to be posted for any forward trades which helps with the default risk, similar to other commodity markets. Users can enter the delivery timing details as they create a request, and dealers will price accordingly.
Most of the hardware has super long wait times right now (16-24 weeks), and there is no market structure today that allows to price the differential between a spot, immediately available B300 server vs one with 16-24 week lead time. Our marketplace allows to observe that.
The market is so new and growing that we sometimes end up working together with competitors to satisfy some of the requests. Would be great to know who it is!
We actually had many rental marketplaces that asked to work with us and validated the need for a hardware marketplace.
glad you like it! we aggregate de-identified pricing from transactions we observe, and whenever you create a quote, you will see an indicative range of what we think your config is worth. This can deviate from the price you get based on the liquidity in the market, but from our experience, it has been super helpful especially for end users.
Anyone we onboard to the marketplace goes through strict know-your-business checks. We never release any payment to the sellers before a transaction is inspected and verified by the buyer, essentially like an escrow service but for free (instead of around 7% at escrow.com).
Our primary focus is on trading the physical hardware, most of the marketplaces we see are rental focused. Our thesis is as open source models become stronger and stronger, owning your own hardware will become more prevalent
This is super cool!
I agree the non-fungible nature of a H100 (or whatever) is a massive challenge. How do you actually verify usage/thermal history on these? Hours run and thermal violations aren't stored on the card IIRC, only ECC error counts and retired pages persist in the GPU's own InfoROM, host side monitoring like DCGM logs, are only as good as whatever the seller hands over. Is condition at inspection based on a real monitoring export from the deployment or just an InfoROM/diagnostic check at the time, and if it's the latter isn't hours run and thermal history basically unverifiable unless we assume good faith and the seller was logging the whole time ?
Essentially how do you avoid having the 'used car problem' without leaning heavily on seller reputation & warranties?
Great question. We don’t treat “used” as one bucket. Before the RFQ goes out, the buyer says what logs and tests they need. If the seller can’t provide usage history, that’s stated in the quote rather than inferred. The market then prices that difference. Hardware with better records and more thorough testing should attract higher bids.
I assume this is a strictly B2B proposition in spite of "built for every side of the GPU market" -- meaning I'm probably not going to be able to buy/bid on small (e.g. 2-8) lots of used, matched RTX 3090s for example?
For now, we are focusing on institutions, since most of this hardware has strict export controls. We deal with all sizes though, we can do a single card to 1024 servers. Can always reach out to us to see if a specific request would fit in the marketplace or not
I’m very curious about how you’re fending off fraud— just wondering about how a lot of dark web transactions supposedly have escrow, but people waive it and then get burned. Do you have anything beyond KYC/escrow?
We use Stripe Connect to transfer the funds, and it is not a waivable process. In addition to KYC documentation, the team personally meets with any participant we onboard to the marketplace. We also have the ability to manually intervene in transactions if need be, but so far we have not had any transaction issues. It almost reminds me of how people were super skeptical of Uber and Airbnb when they first launched, but human trust goes a long way, sometimes further than we originally anticipate.
One thing that guides how we think about this is: we will never ask you to make the payment before you feel comfortable with the counterparty you are dealing with.
But what about the GPUs that are considered defective for corporate and enterprise use, but perfect for personal use...say like A100s that maybe having ECC issues but I can live with that
We ask any seller to post the condition and testing reports for the hardware they sell, so there is no info asymmetry. As long as there is a willing buyer and a willing seller at the conditions specified, transaction goes through! We deal with A100s all the time.
Nice thing about these chips is that they have built in software that allows you to see the condition. The secondary market also supports the circular economy, most of these servers are proving to be useful longer than what the market initially thought. There is always a case to go with used equipment instead of new ones.
As an individual, will I have access to pricing or is that limited to institutions as well?
For that matter, how transparent is this marketplace? Do I need to put in an RFQ for a specific SKU to get a point-in-time price or do I see the whole marketplace?
You mention $300 million dollars in RFQs, but how much of that resulted in chips and cash trading hands? What stops someone from offering the seller a better price to continue the transaction outside of Stoa?
Great questions. We do not accept individual accounts today. Marketplace access and pricing are limited to institutions.
A buyer can create an RFQ for bare GPUs or servers, specifying required configurations while leaving others open. They can see an indicative price range that updates with market activity, then compare all supplier quotes. Other participants’ activity stays private.
Our RFQ activity came from institutions we worked with directly. GPU transactions can take weeks due to sourcing, KYB, financing, and logistics. We are now onboarding those institutions so more of this activity can move through our marketplace.
We create value by making transactions easier for both parties. We enable buyers to compare quotes in one place, while suppliers receive qualified demand and market intelligence without exposing their bids. Binding quotes and non-circumvention terms also help this process.
Will there be support for lower-end GPUs like RTX 4090s or RTX 6000 Pros in the future?
We already support these!
I anticipate some kind of futures on GPUs based on a number of factors such as the availability for rent, the availability at your marketplace, and the models currently available for self-hosting.
Yes, those types of differences across different providers is what the finance world calls basis risk. As more and more transactions take place, we hope to bring transparency into how different components/specs can change pricing, and make them tradable too.
Basis risk is also a big part of futures markets, how much does spot price differ from forward pricing and why. For physical GPUs, theory would expect a downward sloping forward curve due to the depreciation factor, but from what we have seen recently, sometimes not the case!
I don't get it. How it's different from Vast.ai, Salad Cloud, and RunPod?
Those work with rental agreements on compute, we directly transact the hardware itself. Instead of per hr pricing, we do physical unit prices. We help with the shipment and clearing process too.
Oh! You mean like ebay.
ok, Who will handle customs and other complicated things? you guys include that in clearing process?
Yep, we currently only transact within the US in the marketplace. We handle end user disclosure type checks on the platform. For any international deals, we still support them via our over-the-counter brokerage arm today.
This is about buying hardware, not renting it.
Is pricing spot-only or do you support forward commitments? GPU supply deals seem to want term structure - a lab locking capacity for 6 months cares about a very different price than burst spot. Curious how you handle default risk on the longer terms.
We do both! Solving for the forward curve has very interesting applications in the GPU market, as depreciation is a big factor.
We have credit limits/margins in place to be posted for any forward trades which helps with the default risk, similar to other commodity markets. Users can enter the delivery timing details as they create a request, and dealers will price accordingly.
Most of the hardware has super long wait times right now (16-24 weeks), and there is no market structure today that allows to price the differential between a spot, immediately available B300 server vs one with 16-24 week lead time. Our marketplace allows to observe that.
No AMD?
We have AMD chips too, especially MI350X has been in high demand.
No mention on the website as far as I can see.
Website includes some of the most liquid cards, but we have many more configs in the actual marketplace.
I could have sworn another startup doing this exact same thing was on here two weeks ago.
The market is so new and growing that we sometimes end up working together with competitors to satisfy some of the requests. Would be great to know who it is!
We actually had many rental marketplaces that asked to work with us and validated the need for a hardware marketplace.
this is sick — how does price discovery work? RFQ/orderbook?
also how are you dealing with fraud/recourse?
glad you like it! we aggregate de-identified pricing from transactions we observe, and whenever you create a quote, you will see an indicative range of what we think your config is worth. This can deviate from the price you get based on the liquidity in the market, but from our experience, it has been super helpful especially for end users.
Anyone we onboard to the marketplace goes through strict know-your-business checks. We never release any payment to the sellers before a transaction is inspected and verified by the buyer, essentially like an escrow service but for free (instead of around 7% at escrow.com).
hows this different from vast ? or some of the other players inn the space?
Our primary focus is on trading the physical hardware, most of the marketplaces we see are rental focused. Our thesis is as open source models become stronger and stronger, owning your own hardware will become more prevalent
vast.ai rents. this one sells