Used & Refurbished Data Center GPU Market (2026-2036)
The global Used & Refurbished Data Center GPU Market was valued at USD 3.15 billion in 2025. This market is expected to reach USD 21.55 billion by 2036 from an estimated USD 4.12 billion in 2026, registering a CAGR of 18.0% during the forecast period (2026-2036).
- Published
- Oct 2026
- Pages
- 243
- Format
- PDF + Excel
- Report ID
- MR-2305
- Base year
- 2025
- 2025 · BASELINE
- $3.15B
- 2036
- $21.55B
- CAGR 2026–2036
- 18.0%
2025 baseline · 2026–2036 forecast at 18.0% CAGR · hover a bar for the value
Key highlights
The global Used & Refurbished Data Center GPU Market is projected to reach USD 21.55 billion by 2036, driven by shorter refresh cycles of AI fleets, the growth of inference workloads, the expansion of neoclouds, and the large price gap between secondary and new hardware.
North America is expected to account for the largest market share in 2026, while the Middle East & Africa is projected to register the fastest growth during the forecast period.
AI hardware is being replaced far faster than conventional IT. SK tes reported in May 2026 that GPU-dense AI servers are refreshed in as little as 10 to 18 months, against four to six years for traditional enterprise equipment.
By GPU architecture, NVIDIA Ampere is expected to account for the largest market share in 2026, whereas NVIDIA Blackwell & Later is projected to witness the fastest growth through 2036 as early Blackwell fleets turn over.
Secondary GPUs offer a substantial discount to new units. In June 2026, Servnet UK quoted new H100 cards at USD 25,000 to USD 40,000, refurbished units at USD 21,000 to USD 34,000, and used units at USD 12,000 to USD 22,000.
Lifecycle service providers are scaling rapidly on hyperscale decommissioning work. Iron Mountain's asset lifecycle management revenue reached USD 288 million in the second quarter of 2026, up 88%, and Sims Lifecycle Services reported revenue of AUD 757 million for the year to June 2026, up 77.4%.
Report summary
| Particulars | Details |
|---|---|
| Forecast Period | 2026-2036 |
| Base Year | 2025 |
| Estimated Year | 2026 |
| CAGR (Value) | 18.0% |
| Format | PDF, Excel & Cloud Portal · 243 pages |
| Market Size (Value) in 2026 | USD 4.12 Billion |
| Market Size (Value) in 2036 | USD 21.55 Billion |
| Segments Covered | By GPU Architecture: NVIDIA Ampere, NVIDIA Hopper, NVIDIA Volta, Turing & Earlier, NVIDIA Blackwell & Later (includes Rubin), AMD Instinct, Other Architectures (Ada Lovelace L40S/L4, Intel Gaudi/Data Center GPU Max, Others). By Condition: Refurbished, Used. By Form Factor: Complete GPU Servers & Systems, SXM/OAM Modules & HGX Baseboards, PCIe Cards. By Sales Channel: ITAD & Lifecycle Service Providers, Independent Resellers & Brokers, Direct Fleet Sales, Online Marketplaces & GPU Exchanges. By Application: AI Inference, AI Training & Fine-tuning, HPC & Scientific Computing, Visualization, VDI & Rendering, Other Applications. By End User: Neoclouds & GPU Cloud Providers, Enterprises, AI Startups & Model Developers, Research & Academic Institutions, Government & Defense, Other End Users. |
| Countries Covered | North America: U.S., Canada. Europe: Germany, U.K., France, Netherlands, Nordics, Rest of Europe. Asia-Pacific: China, Japan, India, South Korea, Taiwan, Singapore, Australia, Rest of Asia-Pacific. Latin America: Brazil, Mexico, Rest of Latin America. Middle East & Africa: UAE, Saudi Arabia, Israel, Rest of Middle East & Africa. |
| Key Companies | Iron Mountain Incorporated, Sims Limited, TES-AMM (Singapore) Pte. Ltd. (SK tes), Procurri Corporation Limited, Alta Technologies, Inc., Big Data Supply, Inc., The Compute Exchange Inc., server-parts.eu Kft., exIT Technologies, NewServerLife LLC, BrightStar Systems, LA Micro Group (UK) Limited, and PC Server & Parts. |
Report overview
Segments covered: gpu architecture, condition, form factor, sales channel, application, end user.
The growth of this market is mainly driven by faster refresh cycles of AI GPU fleets that are releasing Hopper and Ampere hardware, the cost advantage and immediate availability of secondary GPUs during shortages of new GPUs and memory, and the rapid growth of inference workloads and neocloud capacity. However, rapid price erosion at each new architecture launch, residual value uncertainty, and concerns over reliability, warranty, and provenance restrain the growth of this market.
Furthermore, certified refurbishment, testing, and grading programs, demand from sovereign AI programs and emerging markets, and the rise of GPU-backed financing that depends on reliable residual values are expected to offer growth opportunities for the stakeholders in this market. However, export control compliance and diversion risk, together with the power, cooling, and facility requirements of rack-scale systems, remain major challenges impacting the growth of this market. Additionally, GPU price indices, compute exchanges, and futures, circular data center programs and hyperscaler ITAD partnerships, and the shift toward server-level and rack-scale resale are prominent trends in this market.
The Used & Refurbished Data Center GPU Market comprises the secondary sale of data center graphics processing units and GPU-based systems that have completed a first period of service and are sold to a second owner. The market includes PCIe accelerator cards, SXM and OAM modules, HGX baseboards, and complete GPU servers and systems across NVIDIA, AMD, Intel, and other architectures, from Volta-era hardware to Blackwell and Rubin. The market value covers the final sale value paid by end buyers, with complete servers counted at their full resale value. New GPUs, gaming and consumer GPUs, cloud GPU rental, ITAD service fees, scrap and parts recovery, dealer-to-dealer trades, and untraceable grey-market flows are excluded. The market ecosystem extends from GPU and server manufacturers and first owners to ITAD and lifecycle service providers, refurbishers, resellers and brokers, GPU marketplaces and price index providers, lenders, regulators, and second owners.
The market is experiencing strong growth as AI fleets turn over at an unprecedented pace. NVIDIA has moved to a yearly architecture cadence, from Hopper to Blackwell, Blackwell Ultra, and Vera Rubin, and said in August 2026 that Vera Rubin was in full production, with Rubin Ultra announced for the second half of 2027 and the Feynman architecture for 2028. The company reported data center revenue of USD 193.7 billion for fiscal 2026. Every GPU in that installed base eventually leaves its first owner, so primary shipments set the long-term supply of the secondary market with a lag of three to six years. Microsoft, for instance, retired its V100-based NCv3 virtual machines on September 30, 2025, more than eight years after the V100 was introduced.
Demand is led by neoclouds, enterprises, AI startups, research institutions, and buyers in regions with limited access to new hardware, most of whom deploy secondary GPUs for inference and fine-tuning, where the latest architecture is not essential. Inference volumes are rising sharply: Google said at I/O in May 2026 that it was processing more than 3.2 quadrillion tokens a month, against 480 trillion in 2025. Neocloud capacity is also expanding quickly. CoreWeave reported second-quarter 2026 revenue of USD 2.58 billion, up 112%, and signed an A100 contract extending into 2029, which shows that Ampere hardware continues to earn revenue years after deployment.
Shortages of new hardware add a time advantage to the price advantage of secondary GPUs. SK hynix said in October 2025 that its DRAM, NAND, and HBM capacity for 2026 was sold out, and Micron has said that it expects to sell out its 2026 HBM capacity. Hyperscalers receive the first allocation of new platforms, leaving smaller buyers with the longest queues, while secondary stock can be delivered immediately. Sims Lifecycle Services cites capital expenditure savings of 40% to 60% against new GPUs for its ValidCore recertified GPU offering.
The market is also becoming more transparent and more financially structured. Compute Exchange launched a secondary marketplace for used and refurbished H100 and A100 GPUs in July 2026, Silicon Data publishes a daily H100 rental index, and CME Group and Silicon Data announced plans to launch futures on H100 and B200 rental indices. At the same time, lenders are accepting GPUs as collateral, as shown by CoreWeave's USD 8.5 billion delayed-draw term loan in March 2026 and Lambda's USD 926 million senior secured term loan in August 2026. Together, these developments point to sustained growth in the secondary GPU market, moderated by price erosion at each architecture launch and by export control obligations that apply to resale as well as to new sales.
Market dynamics
13 factors across 5 forcesFaster GPU Refresh Cycles Releasing Hopper and Ampere Fleets
Faster refresh cycles of AI GPU fleets are a major factor driving the Used & Refurbished Data Center GPU Market. Each new NVIDIA generation lowers the cost per token of the previous one, which pushes fleet owners to replace older hardware sooner. SK tes reported in May 2026 that AI-dense servers are being refreshed in as little as 10 to 18 months, and reseller BuySellRam described cycles of five to seven years in the A100 era, three to five years in the H100 era, and 18 to 36 months in the Blackwell era. Iron Mountain reported that its data center decommissioning revenue more than doubled year on year in both the first and second quarters of 2026, reflecting the volume of hardware leaving first-owner fleets.
Cost Advantage and Immediate Availability During New-GPU and Memory Shortages
Secondary GPUs cost substantially less than new units and can be delivered from stock. Servnet UK quoted refurbished H100 cards at USD 21,000 to USD 34,000 and used cards at USD 12,000 to USD 22,000 in June 2026, against USD 25,000 to USD 40,000 for new units. The cost advantage is wider at the system level, because buyers avoid the full price of current rack-scale platforms. With HBM and DRAM capacity for 2026 sold out at major memory suppliers and the latest platforms allocated first to hyperscalers, secondary hardware is often the fastest route to deployable compute for smaller buyers.
Growth of Inference Workloads and Neocloud Capacity
Inference is the main use of older GPUs, and inference volumes are growing rapidly. Most inference workloads do not need the newest architecture, so Ampere and Hopper GPUs remain economically useful long after they leave frontier training. Neoclouds are the largest buyer group for secondary hardware; Nebius reported second-quarter 2026 revenue of USD 582.3 million, up 454% year on year, and guided 2026 capital expenditure of USD 20 billion to USD 25 billion. Compute Exchange reported in July 2026 that buyer requests on its secondary marketplace range from hundreds to tens of thousands of GPUs.
Table of contents
14 chapters · 148 sections · 243 pages · click to expandSegmental analysis
| Segment | Largest share (2026) | Fastest growth (2026–2036) |
|---|---|---|
| By GPU Architecture | NVIDIA Ampere | NVIDIA Blackwell & Later |
| By Condition | Refurbished | — |
| By Form Factor | Complete GPU Servers & Systems | SXM/OAM Modules & HGX Baseboards |
| By Sales Channel | ITAD & Lifecycle Service Providers | Online Marketplaces & GPU Exchanges |
| By Application | AI Inference | — |
| By End User | Neoclouds & GPU Cloud Providers | Enterprises |
By GPU Architecture
- The NVIDIA Ampere segment is expected to account for the largest share of the market.
- The large share of this segment is mainly due to the large installed base of A100 GPUs being replaced by Hopper and Blackwell, low unit prices, and steady demand for inference and fine-tuning.
- However, the NVIDIA Blackwell & Later segment is projected to register the highest CAGR during the forecast period, driven by the first meaningful resale volumes as early Blackwell fleets turn over.
- With Rubin now in full production, Blackwell is expected to become the largest segment by 2036, while Hopper is expected to overtake Ampere by 2030.
By Condition
- The Refurbished segment is expected to account for the larger market share, owing to buyer demand for tested units with documentation and warranty, the higher price of refurbished units, and the certification programs of large ITAD providers.
- The Refurbished segment is also projected to register the higher CAGR during the forecast period, as neoclouds and enterprises increasingly run production AI workloads on secondary hardware.
By Form Factor
- The Complete GPU Servers & Systems segment is expected to account for the largest market share, driven by the decommissioning of complete HGX servers and the higher value of full systems.
- However, the SXM/OAM Modules & HGX Baseboards segment is projected to register the highest CAGR during the forecast period, driven by the resale of Hopper and Blackwell modules and baseboards, as many second owners buy compute trays rather than complete liquid-cooled racks.
By Sales Channel
- The ITAD & Lifecycle Service Providers segment is expected to account for the largest market share, as hyperscalers sell retired fleets through certified partners for data security, chain of custody, and export screening.
- However, the Online Marketplaces & GPU Exchanges segment is projected to register the highest CAGR during the forecast period, supported by new exchanges, independent condition verification, and improving price transparency.
By Application
- The AI Inference segment is expected to account for the largest market share.
- This segment is also projected to register the highest CAGR during the forecast period, driven by the rapid growth of token volumes and the suitability of older GPUs for serving trained models, as GPUs typically move from frontier training into real-time and batch inference as they age.
By End User
- The Neoclouds & GPU Cloud Providers segment is expected to account for the largest market share, owing to the large number of GPU cloud providers and their need for low-cost capacity.
- However, the Enterprises segment is projected to register the highest CAGR during the forecast period, driven by private AI inference and the lower cost of certified secondary hardware.
Geographic analysis
North America
Largest shareIn 2026, North America is expected to account for the largest share of the global Used & Refurbished Data Center GPU Market. The region's position is supported by the largest hyperscaler and neocloud fleets, the largest cluster of ITAD providers and resellers, and export rules that favor domestic buyers of covered hardware. The U.S. is both the largest source and the largest buyer of secondary GPUs, while Canada's CAD 2 billion Sovereign AI Compute Strategy is expected to support demand from domestic firms.
Asia-Pacific
Asia-Pacific is expected to account for the second-largest share of the market. China remains the largest country market in the region, although its growth is moderated by U.S. rules that keep re-exports of covered GPUs to China under a presumption of denial and by the rising share of domestic accelerators. India is projected to grow fastest in the region, as more than 38,000 GPUs have been onboarded under the IndiaAI Mission and demand for affordable compute rises, while Japan is supported by government-funded AI computing projects and new refurbished AI-server cloud services.
Middle East & Africa
Fastest growthHowever, the Middle East & Africa is projected to register the highest CAGR during the forecast period. In November 2025, the U.S. Department of Commerce approved G42 in the UAE and Humain in Saudi Arabia each to purchase the equivalent of up to 35,000 NVIDIA GB300 chips, and these fleets are expected to build local AI ecosystems that later buy and sell secondary hardware.
Europe
In Europe, Germany is the largest market, while the Nordics are expected to grow fastest; Stargate Norway in Narvik targets 100,000 NVIDIA GPUs by the end of 2026, adding to future secondary supply.
Latin America
In Latin America, Brazil is the largest market, while Mexico is projected to grow faster.
Competitive landscape
The global Used & Refurbished Data Center GPU Market is highly fragmented, with large ITAD and lifecycle service providers, independent resellers and brokers, and new GPU marketplaces each serving a different part of the market. Large ITAD providers handle decommissioning and remarketing for hyperscalers, while independent resellers sell tested GPU hardware to enterprises, startups, and smaller cloud providers. Market participants compete primarily on access to hyperscale and neocloud supply, testing and certification depth, warranty terms, provenance and chain-of-custody records, export compliance, pricing, and speed of delivery.
Leading players are expanding processing capacity, acquiring regional ITAD providers, and launching recertification programs and marketplaces. Iron Mountain acquired Wisetek in Ireland, ACT Group in Australia, and Group ATF in France and Belgium to extend its asset lifecycle management network, Sims Lifecycle Services opened a Circular Center in Eindhoven, and SK tes opened facilities in Virginia and Ireland to process retired hyperscale hardware. Acquisitions, facility expansions, product launches, partnerships, and agreements remain the key strategies adopted by major companies.
The report provides a comprehensive competitive assessment of the leading companies operating in the global Used & Refurbished Data Center GPU Market. The key players profiled in the report include Iron Mountain Incorporated (U.S.), Sims Limited (Australia), TES-AMM (Singapore) Pte. Ltd. (SK tes) (Singapore), Procurri Corporation Limited (Singapore), Alta Technologies, Inc. (U.S.), Big Data Supply, Inc. (U.S.), The Compute Exchange Inc. (U.S.), server-parts.eu Kft. (Hungary), exIT Technologies (U.S.), NewServerLife LLC (U.S.), BrightStar Systems (U.S.), LA Micro Group (UK) Limited (U.K.), and PC Server & Parts (U.S.).
- Iron Mountain Incorporated
- Sims Limited
- TES-AMM (Singapore) Pte. Ltd. (SK tes)
- Procurri Corporation Limited
- Alta Technologies, Inc.
- Big Data Supply, Inc.
- Compute Exchange Inc.
- NewServerLife LLC
- BrightStar Systems
- LA Micro Group (UK) Limited
- PC Server & Parts
Expert perspectives
The Used & Refurbished Data Center GPU Market is shaped less by whether retired GPUs are still useful than by how quickly their value falls. Supply is set to rise sharply as the Hopper and early Blackwell cohorts shipped from 2023 onward reach the end of their first service life, and inference demand gives these GPUs a productive second life. Yet price steps at each architecture launch, uncertain failure rates, and inconsistent grading mean that unit volumes will grow faster than market value, and sellers that hold inventory through a transition carry real risk.
Three structural changes are expected to shape competitive positions through 2036. First, certification is becoming the price of entry: buyers and lenders increasingly pay for stress-tested, serial-documented hardware, which moves value from brokers toward lifecycle providers that can prove quality at scale. Second, the market is financializing, as rental indices, compute futures, and GPU-backed loans create a common reference for residual value. Third, the unit of trade is shifting from cards to servers and liquid-cooled racks, which rewards sellers that can match hardware to buyer facilities.
For companies planning entry or expansion, the most attractive positions over the forecast period are likely to be found in certified recertification and warranty programs, hyperscale decommissioning partnerships, residual value and remarketing services for lenders, and secondary supply to inference-focused neoclouds, enterprises, and sovereign AI buyers in allied markets. The principal risk is compliance: as enforcement against diversion intensifies, sellers without serial-level chain-of-custody records and robust buyer screening will find their access to hyperscale supply narrowing.
Customer perspectives
Insights gathered during primary interviews with infrastructure, procurement, and asset recovery professionals operating in this market highlight where purchasing priorities are shifting. The following perspectives reflect recurring themes raised across these discussions.
This reflects the growing premium placed on certified, documented hardware and the role of inference in extending the economic life of Hopper-generation GPUs.
This indicates enterprise demand for complete, air-cooled secondary systems that can be deployed quickly in existing data center space.
This points to strong demand for compliant, certified ITAD partners and to the timing pressure that price erosion places on fleet disposal decisions.
“For inference, an H100 that is two years old still earns its keep. What we will not buy any more is an unknown card. We want burn-in reports, thermal history, and a serial-level record of where it has been, and we will pay more for refurbished units that come with that paperwork and a real warranty.”
“New allocation for current platforms is months away for a company our size. Refurbished 8-GPU servers let us stand up private inference this quarter, inside our own facility, without redesigning power and cooling for a liquid-cooled rack.”
“Every disposal now goes through export classification and buyer screening before anything leaves the building. We only work with partners who can show chain of custody to the serial number. Speed matters, because values drop quickly once the next generation ships.”
Frequently asked questions
The global Used & Refurbished Data Center GPU Market is estimated at USD 4.12 billion in 2026.
Cite this report
Meticulous Research. (2026). Used & Refurbished Data Center GPU Market - Global Opportunity Analysis and Industry Forecast (2026-2036) (Report No. MR-2305). Meticulous Market Research Pvt. Ltd. https://www.meticulousresearch.com/product/used-refurbished-data-center-gpu-market-6988