H200 vs H100 in 2026: memory, China licensing and the used market

On 13 January 2026 the US Bureau of Industry and Security said it would review licence applications for exports of the NVIDIA H200, the AMD MI325X and similar chips to China on a case-by-case basis, provided certain security requirements were met, replacing a presumption of denial. The change followed President Trump’s announcement on 8 December 2025 that approved Chinese customers could receive the H200. For buyers elsewhere, the significance is that a GPU many had treated as the end of a product line acquired a large new market.

What the H200 adds over the H100

NVIDIA’s specification page lists the H200 SXM with 141 GB of HBM3e and 4.8 TB/s of memory bandwidth, which NVIDIA describes as nearly double the capacity of the H100 and 1.4 times the bandwidth. Its inference claims, such as up to 1.9 times H100 throughput on Llama 2 70B, are memory-driven: both are Hopper GPUs. The PCIe H200 NVL is a dual-slot, air-cooled card with the same 141 GB, a configurable TDP of up to 600 W and 2- or 4-way NVLink bridges at 900 GB/s per GPU; NVIDIA puts its memory at 1.5 times that of the H100 NVL. For workloads that fit comfortably in 80 GB, the H100 remains a capable card. For larger models and long context windows, the H200 avoids splitting a model across more GPUs than necessary.

Availability has been shaped by China. The Register, citing Reuters, reported on 31 December 2025 that Chinese companies had placed orders for more than two million H200 chips, that NVIDIA held about 700,000 in stock and had approached TSMC about producing more. An NVIDIA spokesperson said the company did not expect China sales to affect supply for US customers. Beijing’s approval was still uncertain at the time.

The used and refurbished market is where the H100 is now most visible. Servnet, a UK reseller, published indicative ranges in June 2026 of roughly $12,000 to $22,000 for used H100 cards and $21,000 to $34,000 for refurbished units, and said H200s were increasingly available second-hand without quoting prices. These are one vendor’s figures, and it advised confirming them on the day. Its checklist is a sensible baseline: provenance from enterprise or hyperscaler fleets, a warranty and RMA path, firmware state and a burn-in or health report.

What it means for buyers

Decide on memory first: if the model and context do not fit in 80 GB per GPU, compare H200 against Blackwell rather than against H100. Specify the form factor precisely, because H100 PCIe, H100 NVL, H100 SXM, H200 NVL and H200 SXM are different products with different servers, power and bridge requirements. For new H200 supply, ask for written lead times and check whether the offer depends on stock that may also be sought for China. For used units, request serial numbers, the NVIDIA warranty status, firmware versions and test logs before payment. Export licensing rules apply to Hopper parts just as they do to Blackwell, so the end-user statement belongs in the first exchange, not the last.