AI accelerators and GPUs (Nvidia H100, A100, and the export-tuned H20)

AI accelerators and GPUs are specialised processors that perform the massively parallel computation used in modern artificial-intelligence training and inference. Nvidia's A100, H100 and export-tuned H20 illustrate how product performance, software support, memory bandwidth, interconnection and export classification combine to shape access to advanced computing. The devices are dual-use technologies; their strategic significance depends on the systems, users and policies around them.

Function

An AI accelerator combines many parallel arithmetic units with high-bandwidth memory. The H100, launched in 2022, pairs Hopper-generation compute with HBM3 delivering about 3 terabytes per second of memory bandwidth in listed configurations. Large training workloads use multiple devices linked by high-speed interconnects, so useful capability depends on chip count, networking, memory, power, cooling and software. Nvidia's CUDA ecosystem is an important part of that system, while fabrication and packaging add further supply dependencies. Cluster-scale requirements are treated at Compute clusters and AI datacentres (sovereign AI compute).

Product names do not denote a fixed capability across every board or system. Memory capacity, interconnect configuration, power envelope and software optimisation affect performance, and published specifications describe vendor configurations rather than the output of a deployed cluster.

Control and weaponisation history

The October 2022 controls used performance and interconnect thresholds that covered advanced processors including listed A100 and H100 configurations. Later rules revised the thresholds and licensing structure, and suppliers developed products including the H20 for markets subject to controls. The January 2025 AI diffusion framework added destination groups and country allocations, while Commerce announced categorical non-enforcement and planned rescission in May 2025 without completing formal removal. On 13 January 2026, BIS announced case-by-case licence review for H200, AMD MI325X and similar chips subject to stated conditions. The framework's codified but non-enforced status is treated at United States AI diffusion rule, non-enforcement and destination-specific controls (2025-present). Classification and licensing remain product, destination, end-user and end-use questions under the current BIS rules.

Contestation

Whether processor controls can materially restrict AI capability is contested. Hardware restrictions can raise acquisition costs and delay cluster formation, but substitution, stockpiling, cloud access and trans-shipment can reduce their effect. Tighter controls may also encourage domestic alternatives, the self-undermining arsenal dynamic. A second design question is whether capability can instead move through model weights, distillation or hosted services. Proposals for hardware-enforced governance are treated at Compute governance and on-chip location/verification mechanisms. None of these pathways makes chip performance alone a measure of national capability.

Current status and evidentiary limits

A100, H100 and the export-tuned H20 are product generations, not export-control categories. Legal treatment depends on the applicable ECCN thresholds, destination, end user and end use. BIS announced on 13 January 2026 that H200, AMD MI325X and similar chips could receive case-by-case licence review subject to conditions. As at 30 July 2026, current part 742 and related EAR provisions governed classification and licensing. Product performance or marketing does not prove military use, licence status or state intent.

See also

AI-chip and compute export control · Bureau of Industry and Security (US Commerce) · United States advanced-computing and semiconductor controls on China (2022-present) · Compute clusters and AI datacentres (sovereign AI compute) · Economic statecraft

Sources

• Nvidia H100.
• Nvidia A100.
• EAR part 742.
• BIS 13 January 2026 advanced-computing policy.
• GAO B-337935.