Your spreadsheet has a line item with 64 H100s and a 4-month old price quote.
How much are they worth now?
Most finance teams price AI hardware in the simplest terms: acquisition cost minus depreciation over time. GPU value doesn’t track like that. Instead of a slow downward slope, it’s a price chart with sudden drops in between periods of stability.
NVIDIA’s announcement of Blackwell caused one of those sudden drops. You need to know what this announcement changes and what it doesn’t before you lock in a deal for your used GPUs.
The Market Is Not Repricing One GPU
It’s easy to fall into the trap of talking about a fleet of H100s like they’re all the same asset.
An eight-GPU DGX H100 is not the same thing as a loose pile of PCIe cards. The market discounts hardware based on buyer’s deployment speed, validated topology, and usable memory footprint.
In 2024, NVIDIA announced Blackwell architecture as the successor to Hopper (H100 and H200 models), marking a leap forward in efficiency. According to company documentation, Blackwell cuts large-language-model inference cost by up to 25x compared to Hopper.
Less than 12 months later, Google Cloud was already previewing A4 instances based on NVIDIA HGX B200. Google’s own blog post says that each B200 GPU delivered 2.25 times the peak compute and 2.25 times the HBM capacity of the previous-generation A3 High H100 platform.
| Deployment Speed | Validated Topology | Usable Memory Footprint |
|---|---|---|
| Buyers pay for hardware that is ready to run. They don’t care about silicon that requires integration work before it earns its first inference call. | A known, documented network fabric and host BOM is a premium asset. | Memory capacity and bandwidth determine which workloads a system can actually run, and how it competes against Blackwell-class alternatives. |
The Data Says Why H100 Value Compresses
This table shows the real reason why H100 value dropped so quickly.
| Platform | Current official signal | What that means for resale |
| H100 | NVIDIA still positions H100 as a data-center GPU for training, inference, and scale-out workloads | The silicon is not obsolete just because it is no longer the newest badge |
| H200 | NVIDIA says H200 offers 141 GB of HBM3e at 4.8 TB/s, nearly double H100 capacity with 1.4x more memory bandwidth | H100 gets pressured first on memory-heavy inference and larger-context workloads |
| Google A3 H100 | Google still documents A3 Mega, A3 High, and A3 Edge H100 instances for pre-training, fine-tuning, and serving inference | H100 remains commercially useful in current cloud catalogs |
| Google A4 B200 | Google says A4 B200 delivers 2.25x peak compute and 2.25x HBM capacity per GPU over A3 High H100 | Buyers now have a much stronger next-gen alternative when comparing full-system purchases |
| DGX H100 | NVIDIA DGX H100 ships with 8x H100 GPUs, 640 GB total GPU memory, and 7.2 TB/s aggregate GPU interconnect bandwidth | A complete DGX H100 is still a real system, not just a component lot |
| DGX B200 | NVIDIA says DGX B200 delivers 3x DGX H100 training performance and 15x DGX H100 inference performance, with 1,440 GB total GPU memory and 14.4 TB/s aggregate NVLink bandwidth | The premium for an older turnkey AI box compresses fast once the new turnkey box exists |
Your H100 isn’t the default GPU anymore, but it still has a defined market with enterprise use cases. The more your valuation depends on “top of stack” scarcity, the faster it gets repriced. It’s worth comparing how A100 and H100 stack up before making any decisions.
What Hardware Depreciates First?
The first thing to lose value is not necessarily the intact H100 unit. Incomplete systems get punished before anything else first.
| 01 Incomplete Systems | 02 Turnkey Premium on Older Integrated Systems | 03 Hard-to-Support Inventory |
|---|---|---|
| If you break an HGX H100 node into parts, lose the network fabric context, separate the trays from the host, or force the next buyer to reverse-engineer exactly what they are getting, you have already traded a deployment asset for a parts problem. Once Blackwell-class alternatives are in the market, buyers become less patient with parts problems. | NVIDIA’s own DGX numbers make that unavoidable. DGX B200 does not just edge past DGX H100. NVIDIA markets it as a materially different performance class with far more memory and interconnect bandwidth. That means a buyer considering a full-box acquisition is no longer deciding between “buy H100 or wait.” They are deciding how much discount they need before taking the older box instead of the newer standard. | A clean DGX H100 with a documented configuration, complete inventory, known health, and serviceable condition is one market. A mixed lot of H100-era accelerators with weak provenance, missing networking, or uncertain runtime history is another. When the market gets more options, uncertainty gets repriced faster than performance. |
What Holds Value Longer Than You Might Think
H100 owners don’t need to panic. The market has not stopped using it.
Google’s current accelerator-optimized machine documentation still lists H100-based A3 Mega, A3 High, and A3 Edge systems alongside H200-based A3 Ultra and B200-based A4 offerings. They’re still in production, so they’ll be attractive for customers that want strong performance without paying the newest-generation premium.
Complete, deployable H100 systems still matter. You can sell Nvidia DGX servers and AI GPUs as complete systems and recover meaningfully more than you would by breaking them apart.
| What Holds Value Best 1. Complete DGX H100 systems with intact configuration and clean provenance 2. Complete HGX H100 nodes with the networking and host context buyers actually need 3. H100 inventory already qualified for real inference or training workloads 4. Lots large enough to solve a buyer’s capacity problem in one transaction | What Holds Value Worst What holds value worst is anything that makes the next owner go under the hood. If your hardware forces the buyer to figure out missing BOM details, uncertain firmware state, unclear topology, or whether the lot is really complete, you are no longer selling AI infrastructure. You’re selling a project that may not be worth their time. |
Hitting Your Finance Goals Is A Matter of Timing
Business leaders usually want a single number for all their assets. The better way to look at it is a pricing curve based on release timing.
Once a newer generation like Blackwell becomes visible in cloud previews, current platform docs, and integrated system catalogs, the discount for waiting compounds against you.
That is especially true if your current fleet sits in the middle.
| The Top End This still moves because buyers want deployable capacity fast. Complete, well-documented systems with clean provenance continue to attract bids from buyers who need capacity now and cannot wait for Blackwell availability. | The Middle Gets squeezed first. This is where the buyer starts asking whether they should just stretch to H200 or B200 instead. The middle is the most dangerous place to sit when a new generation becomes the reference point. | The Bottom End This still clears because customers want parts and lower-tier use cases exist. But the floor is lower than it was, and the discount for uncertainty compounds faster as Blackwell-class supply grows. |
If you are holding H100, HGX H100, or DGX H100 inventory, delaying a sale can cut into your return.
How To Adjust Your Refresh Plan Accordingly
If your team expects to retire H100-era gear, the goal is not to predict a magic resale number from a spreadsheet.
The goal is to preserve the part of the asset the buyer will still pay for.
| 1. Document as System | 2. Separate Premium Inventory | 3. Record Full Configuration | 4. Package for Market |
You’ll need to document the hardware as a system before it gets broken apart. Separate premium inventory from tail inventory before the de-rack turns everything into a mixed lot. Record exactly what you have: GPU count, form factor, host platform, networking, storage, cooling context, and any deployment details. These details will make you a better partner with your buyer, who wants to know if the lot is immediately useful.
It also means being honest about what you are selling.
If the fleet is really a set of full DGX H100 systems, sell it like a full DGX H100 fleet.
If it’s loose boards and leftover nodes, make sure everyone understands that. The buyer and seller both need to know exactly what’s changing hands.
The companies that recover the most value from AI hardware need more than a detailed spreadsheet model. Companies that lock in deals before the project gets messy and package the asset in the shape the market still wants will consistently succeed.
Dutch 
English 
German 
Spanish 
French 
Swedish