Compute Exchange launches secondary marketplace for used Nvidia H100 and A100 GPUs
Compute Exchange has launched a secondary marketplace for used and refurbished Nvidia H100 and A100 GPUs, aimed at enterprises and cloud providers seeking to scale AI infrastructure at lower costs. The platform provides a service for sourcing physical hardware, complementing its existing GPU capacity procurement business.
Key Takeaways
- The new marketplace focuses on H100 and A100 generations, which the company identifies as a primary concentration of secondary market demand.
- Hardware requests on the platform have scaled from hundreds to tens of thousands of GPUs across startups and infrastructure operators.
- Compute Exchange maintains a market-neutral position, managing sourcing, logistics, and transaction execution between buyers and suppliers.
- Silicon Data Inc., led by Compute Exchange CEO Carmen Li, provides the real-time GPU pricing benchmarks used to value the marketplace's inventory.
Why It Matters
This move signals a maturing AI hardware lifecycle where older silicon remains viable for production workloads rather than facing immediate obsolescence. For the streaming and video infrastructure ecosystem, secondary markets provide a critical cost-savings lever as generative AI and intensive transcoding tasks drive up compute budgets; utilizing refurbished H100s can significantly lower the barrier to entry for localized or specialized AI clusters. The development suggests that hardware availability is shifting from a pure scarcity model to a structured financial asset class with transparent liquidity. Watch for the impact of Nvidia's Blackwell rollout on secondary H100 pricing, which could further drive down CapEx for mid-tier cloud providers by late 2026.
Additional Context
The launch of a formal secondary market arrives as the GPU landscape undergoes a significant pricing correction. Per secondary market trackers in early 2026, used H100 units that commanded premiums during peak scarcity have seen valuations stabilize between $25,000 and $40,000, while refurbished A100 80GB variants frequently trade in the $15,000 to $17,000 range. This price discovery is increasingly critical; as noted by Substack's analysis in May 2026, compute is effectively becoming a new asset class, illustrated by CME Group's partnership with Silicon Data to explore futures markets for computing power. Simultaneously, the supply-demand dynamic is shifting due to the mass delivery of Nvidia’s Blackwell (B200) architecture. According to reports from Medium in June 2026, large-scale upgrades to Blackwell systems are expected to flood the secondary market with previous-generation Hopper hardware, putting further downward pressure on the H100 and H200 prices. While Blackwell offers significant performance gains for models exceeding 70 billion parameters, industry data suggests H100s remain highly cost-competitive for standard inference tasks, frequently running at roughly $1.80 to $3.50 per hour on specialized clouds compared to $7.00 per hour for newer alternatives. Enterprise procurement strategies are also adapting to these hardware cycles. Beroe Inc. reported in April 2026 that infrastructure certainty has become as important as cost optimization, with many firms moving toward 18-to-24-month ownership models for high-utilization workloads. As lead times for new Nvidia hardware shortened to 6-12 weeks in early 2026—down from the 50-week peaks seen in 2023—the availability of immediate secondary inventory provides a critical buffer for organizations facing electricity and data center permitting delays.
Read full article at siliconangle.com
Get this in your inbox → Subscribe
Enjoy our coverage?
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source