| DeepSeek-V3 efficiency | DemonstratedA 671B-parameter mixture-of-experts model delivered strong author-reported benchmark results; on METR’s autonomy suite it was comparable to Claude 3.5 Sonnet (Old) while trailing newer frontier models. | MaterialThe disclosed 2.788 million GPU-hours exclude prior research, ablations, data, staff, acquisition, energy, and failed work. Extensive low-level co-design was required. | HighTraining used Nvidia H800 accelerators and a software/networking stack centered on controlled U.S. technology. | MixedAlgorithms and weights diffuse readily; installed chips cannot be recalled. Future hardware refresh, HBM, and accelerator replenishment remain enforceable chokepoints. | Adaptation + cost penaltyModerate confidence |
|---|
| Pangu domestic training | DemonstratedHuawei reports training a 718B-parameter model from scratch on Ascend and says its system supports all training stages, with self-reported results near DeepSeek-R1 on several tasks. | HighThe run used 6,000 accelerators, about 19 trillion tokens, 30 percent model FLOP utilization, and extensive recomputation, swapping, topology, operator, and simulation work. | ReducedAscend and Huawei’s CANN/MindSpeed layers are domestic, but the disclosed stack also uses open-source PyTorch, Transformers, and NVIDIA-origin Megatron-LM. Public evidence does not resolve HBM, fabrication equipment, packaging, optics, or the provenance of every die. | MixedDomestic accelerators reduce leverage at the finished-chip layer. HBM, EDA, fab tools, foundry services, and packaging remain alternative control points. | Genuine weakeningModerate confidence |
|---|
| CloudMatrix systems engineering | DemonstratedThe vendor-authored evaluation serves a 671B-parameter DeepSeek-R1 model with little reported INT8 accuracy loss across 16 benchmarks. | HighThe complete system uses 384 NPUs, 192 CPUs, a large optical fabric, disaggregated memory, custom collectives, scheduling, quantization, and fault recovery. | ReducedHuawei designed the accelerator, CPU, and CANN layers, while the disclosed stack interoperates with PyTorch, TensorFlow, ONNX, and Kubernetes; HBM, fab tools, packaging, and high-speed optics remain unresolved. | MixedPooling weakens a per-chip threshold. Aggregate HBM, networking, energy, manufacturing equipment, and production volume remain observable constraints. | Adaptation + cost penaltyModerate confidence |
|---|
| Rerouted H100/H200 access | UnknownThe chips could support advanced training and inference, but public court and agency materials do not identify a completed model or workload. | MaterialThe route required straw purchasers, intermediaries, mislabeled hardware, false paperwork, warehouses, logistics companies, and large cross-border payments. | CompleteThe route depended on Nvidia accelerators and foreign procurement and logistics. It created no domestic technological independence. | LowCustomer checks, beneficial-ownership review, shipment inspection, data-center verification, and financial tracing can disrupt this route; this case was disrupted. | Circumvention, not independenceModerate confidence |
|---|
| State support evidence gap | UnlinkedNational tax preferences and local compute vouchers support the sector, but no reviewed award record ties them to a specific restored capability in this MVP. | UnknownSubsidies may transfer cost to the state without removing real hardware, energy, or time requirements. | UnknownA tax preference or compute voucher does not disclose the origin of chips, memory, networking, or fabrication inputs. | IndirectFiscal support can absorb higher prices but cannot itself replace a denied component or protect a foreign access route. | Insufficient evidenceLow confidence |
|---|