Where will a batch of H100s purchased today end up in a few years?
For cloud providers or AI infrastructure operators with tens of thousands of GPUs, the introduction of next-generation GPUs into data centers does not mean the original equipment immediately loses its value. They may continue to run, be reassigned to other workloads, relocated to other data centers, or sold to new users.
Thus, a very practical question begins to emerge:
Will data center GPUs, like servers, automobiles, or large industrial equipment, eventually form a scaled secondary market?
The current answer is: Used GPUs have entered resale channels, but there is still a significant distance to go before a truly mature, large-scale secondary market is established.
What truly determines whether this market can scale is not just how many old GPUs will exist in the future, but whether these devices can continuously enter the market, and whether there are enough "second owners" willing to take them over and generate economic returns from them.
The full analysis text is below; for the video version, please visit: https://weixin.qq.com/sph/Apuhm3UUP
Used GPUs Already Have Trading Channels, But Are Far from a Mature Market
Currently, GPU trading matchmaking platforms primarily serving enterprise customers and data center operators have emerged in the market. Taking Compute Exchange as an example, the platform provides matchmaking services for GPUs such as H100, H200, A100, L40S, and V100. The trading targets include used equipment, refurbished equipment, and OEM surplus inventory, while also providing reference quotes for different models.
This indicates that data center GPUs already have relatively professional distribution channels. However, it is too early to say that a mature secondary market has been formed.
The most direct issue is that publicly available information on actual transactions remains limited, and the market has not yet formed a transparent and stable pricing system.
The price ranges published by Compute Exchange are mainly compiled based on supplier quotes and brokerage market information, serving only as references and not representing the final transaction prices of the equipment. For specific transactions, prices are also affected by factors such as purchase quantity, region, equipment condition, specific configurations, and warranty terms. The platform's public pages also do not disclose specific transaction prices, transaction volumes, and complete transaction records.
Therefore, although quotes for used H100s can be seen today, there are still many key questions lacking clear answers regarding equipment pricing and transaction efficiency:
For H100 servers of the same model and configuration, with similar usage years and equipment conditions, what range do the actual transaction prices usually fall into?
What should the price difference be for similar equipment in different regions?
How much price difference do runtime, equipment health status, and remaining warranty respectively bring?
For equipment of different models and quantities, how long is the actual transaction cycle usually?
Only when equipment with similar configurations and conditions can form a relatively stable and comparable transaction price range, and asset holders can roughly estimate the sales cycle, will the secondary GPU market have the foundation for further scaling.
A More Critical Question Is: When Will the Supply of Used GPUs Surge?
Price opacity is only one aspect of the current market's immaturity. Another more critical issue is: existing public information is not yet sufficient to prove that the equipment entering the secondary market has formed a continuous, stable, and large-scale supply.
The H100 entered full-scale production in September 2022. The first batch of boards and systems were delivered successively in the autumn of that year, and cloud service providers such as AWS, Google Cloud, Microsoft Azure, and Oracle began deploying H100 instances in 2023. By March 2023, the DGX H100 entered full-scale production, further expanding the supply of H100 in cloud services and complete machine systems.
In other words, the time that large batches of H100s have actually been put into use is not long, and they have not yet generally entered the stage of centralized disposal.
More importantly, the launch of next-generation GPUs does not mean that previous-generation equipment will quickly enter the secondary market.
For cloud providers and data center operators, even if a batch of H100s is no longer the latest product, as long as they can still handle inference, training, or other computing tasks at a reasonable cost, they may continue to be used in the original clusters. Operators can also reassign them to other workloads or relocate them to data centers with lower costs.
Therefore, there is no simple correspondence between technology upgrades and the supply of used equipment.
What truly determines whether a batch of GPUs enters the secondary market is how operators choose among continuing operation, internal reassignment, relocation, and sale. Only when selling becomes more attractive will an increasing number of devices truly flow into the external market.
This also means that to judge whether the secondary GPU market can significantly scale up, one cannot only look at the deployment scale of new platforms like Blackwell and Rubin, but also whether early deployed equipment such as H100s has begun to be continuously and in batches sold by large holders.
Why Are Used GPUs Harder to Price
Even if more equipment enters the secondary market in the future, pricing used GPUs will not be as simple as pricing ordinary consumer electronics.
First, although all called "H100," the actual items traded may be completely different. Some are PCIe accelerator cards, some are SXM modules, some are already installed in HGX servers, or even sold as part of a complete rack system. Different form factors mean different supporting network, power, and cooling conditions, as well as different costs for disassembly, transportation, and redeployment.
Even if the model and configuration are the same, the condition of the equipment itself will also affect the price. How long the equipment has accumulated runtime, what temperature environment it has been in for a long time, whether ECC errors have occurred, its physical condition, and how long the remaining warranty period is will all affect the buyer's judgment.
Therefore, when making large-scale purchases of used GPUs, buyers usually cannot just look at the model and quote. The verification items listed by Compute Exchange include runtime, temperature records, diagnostic information from the NVIDIA System Management Interface, and ECC error records; for large purchases, buyers and sellers usually also agree in advance on stress tests or return periods.
Currently, some refurbishers and IT asset disposition providers already offer services such as diagnostics, stress testing, cleaning, and component replacement. However, from public market information, a grading standard universally adopted by buyers and sellers that can directly map test results to transaction prices has not yet been seen. Compute Exchange itself is only responsible for verifying counterparties and matchmaking buyers and sellers, and is not responsible for testing, certifying, or guaranteeing equipment conditions.
This brings up a very practical issue: for the same batch of two-year-old H100s, different buyers may have vastly different judgments regarding equipment condition and risk.
If buyers cannot accurately judge the equipment condition, they often demand a larger discount to cover the risk; if sellers think the discount is too high, they may be unwilling to sell. The harder it is to compare equipment conditions, the harder it is for buyers and sellers to quickly reach a consensus on price.
Therefore, to further expand the secondary GPU market, in addition to needing more buyers and sellers, a more mature testing, grading, and valuation system is also required.
Only when equipment conditions can be described according to relatively unified standards and clearly reflected in prices will it be easier for both parties to reach an agreement.
What Truly Determines Whether the Market Can Scale Up Is the Presence of "Second Owners"
Even if large cloud providers begin to continuously sell previous-generation GPUs in the future, it does not mean the secondary market will definitely expand rapidly. How much equipment is put up for sale only answers the supply-side question; more importantly, are there enough buyers willing to take them over.
For potential buyers, the lower price of used GPUs is certainly attractive, but what really needs to be considered is whether it is cost-effective after purchase.
The procurement price is only part of the cost. Compared with next-generation equipment, previous-generation GPUs may be at a disadvantage in terms of computing efficiency, energy efficiency, and system density under certain workloads. To provide the same scale of computing power or complete equivalent workloads, buyers may need to deploy more equipment, occupy more rack space, and bear higher power and maintenance costs. In this way, the advantage brought by low-price procurement will be weakened.
If it also involves server and network reconfiguration, equipment transportation, installation, and business migration, the overall cost will further increase.
Therefore, whether used GPUs can form stable demand ultimately depends on a very practical economic calculation: can the revenue or business value brought by these equipment cover the costs of procurement, data center space, power, migration, and O&M, and leave a reasonable return for buyers.
This also determines which users are more likely to become the main buyers of used GPUs. They may not need the most advanced chips, but they need workloads suitable for previous-generation GPUs, along with the capability to redeploy and continuously operate these devices.
If the number of such buyers is limited, even if a large number of used GPUs appear on the market, it may only be an increase in supply, rather than the formation of a truly scaled secondary market.
Therefore, the key to determining whether the secondary GPU market can scale up is not how much old equipment will be sold in the future, but whether there are enough "second owners" who can make the economic calculation work after purchasing these devices.
Under What Circumstances Can the Secondary GPU Market Be Considered Truly Scaled?
Rather than judging now in which year the secondary GPU market will scale up, it is better to look at several more practical changes.
First is whether the supply can appear continuously. If the market only has occasional scattered batches, it is difficult to support a scaled secondary market. What is truly worth paying attention to is whether large cloud providers, GPU cloud service providers, and data center operators have begun to continuously sell previous-generation equipment and gradually form a stable source of supply.
With supply, we also need to see whether there are enough buyers to absorb it. After used equipment enters the market, if transactions can be completed within a relatively stable and predictable cycle, it indicates that the market has begun to form relatively stable absorption demand.
At the same time, the equipment itself needs to be easier to compare. Information such as runtime, health status, fault records, remaining warranty, and specific configurations should ideally be described according to relatively unified standards and reflected in the price. Otherwise, even for equipment of the same model, it is difficult for buyers and sellers to judge a reasonable price.
Another important indicator is that public, verifiable actual transaction information gradually increases. Knowing only "how much a certain GPU is quoted" is not enough; the market also needs to know at what price equipment with similar configurations and conditions actually transacts. As such data continues to accumulate, the price ranges corresponding to different models, configurations, and equipment conditions will become clearer.
Finally, it also depends on whether these devices can be redeployed at a reasonable cost. Disassembly, transportation, installation, data center renovation, and software adaptation all require investment. If the redeployment cost is too high, even if the equipment itself is cheap, it may not be cost-effective for buyers.
Therefore, a truly scaled secondary GPU market cannot rely solely on "more and more old equipment." It requires continuous supply, stable absorption demand, relatively unified equipment grading, increasingly transparent actual transaction prices, and acceptable redeployment costs.
When these conditions are gradually met, secondary GPU transactions can develop from scattered, occasional trades into a market with larger scale, more transparent prices, and more predictable transaction cycles.
2027—2029 May Be a Window Worth Watching
Currently, there is not enough basis to judge in which year the secondary GPU market will significantly scale up. However, looking at the deployment time of the H100, 2027—2029 may be a window worth close observation.
The H100 began delivery in the second half of 2022, and its supply in cloud services and complete machine systems further expanded in 2023. By 2027—2029, it will have been about four to seven years since the H100 began delivery and expanded deployment in 2022—2023. Meanwhile, new platforms like Blackwell continue to be deployed, and Rubin has also entered the full-scale production ramp-up phase, with related mass-produced systems planned to be shipped successively starting in the autumn of 2026.
This means that by 2027—2029, the market may more easily see clearly how the first batch of large-scale deployed H100s will be subsequently used and disposed of.
At that time, there are two signals most worth paying attention to.
First is whether the supply side has changed: whether large cloud providers, GPU cloud service providers, and data center operators have begun to continuously and in batches sell earlier deployed equipment such as H100s, rather than only having scattered batches enter the market.
Second is whether these devices can be smoothly absorbed by the market: whether H100s entering the secondary market can find new buyers within a relatively stable and predictable cycle, and whether public transaction information begins to increase.
Only if these two signals appear simultaneously does it mean that the secondary GPU market may begin to move from scattered transactions to scaling up.
By then, the most worthwhile question to answer is: of the first batch of large-scale deployed H100s, how many have begun to enter the secondary market, and how many have truly found "second owners"?
There is currently no definitive answer as to when the data center GPU secondary market will truly scale up.
What can be confirmed is that used equipment such as H100s and A100s has entered professional matchmaking and resale channels, but whether the market can form continuous supply, whether there are sufficiently stable buyers, whether equipment grading standards can be gradually unified, and whether actual transaction prices can become more transparent, all remain to be seen.
How large this market can ultimately become depends not only on the iteration and deployment of next-generation GPUs, but more on whether the first batch of large-scale deployed data center GPUs will continuously enter the secondary market, and whether they can find enough "second owners" after entering the market.
If such buyers can continuously emerge and form stable absorption demand, the data center GPU secondary market will have the possibility to truly scale up.
By then, what the industry cares about will not just be: how fast is the next-generation GPU?
There will also be another increasingly practical question: how much is the previous-generation GPU still worth?