China's AI industry shows a split scorecard: on models, DeepSeek-V3's open-weight, MIT-licensed mixture-of-experts architecture and DeepSeek-R1's benchmark parity with OpenAI's o1 preceded a $593 billion single-day NVIDIA sell-off, while the widely cited $5.576 million cost figure covers only the final training run. On hardware, Huawei's Ascend 910C still delivers about 60% of NVIDIA H100's per-chip performance, and HBM supply constraints cap actual 2025 shipments far below estimated capacity, even as US export controls kept tightening through January 2026.
How has US export control escalation pushed China toward open-source breakthroughs?
Washington tightened chip export rules on China's AI sector three separate times within about 21 months. NVIDIA's China-specific H20 chip has required an export license since April 9, 2025, effectively halting unlicensed salesCITE:E7. On May 13, 2025, the US rescinded its earlier "AI Diffusion Rule" and warned that using Huawei Ascend chips could violate export controlsCITE:E7. Most recently, on January 13, 2026, the policy shifted again to case-by-case license review for chips including the H200CITE:E7. Separately, a claim circulating that Chinese firms must pay 15% of related China revenue to the US government is not backed by any published BIS rule text and remains unverifiedCITE:E7.
How do DeepSeek-V3's MoE architecture and open-weight release achieve frontier-level capability?
DeepSeek released DeepSeek-V3 as an open-weight model under an MIT license, permitting commercial useCITE:E1. The model uses a mixture-of-experts (MoE) architecture with 671 billion total parameters, of which only 37 billion are activated per token, and it was trained on 14.8 trillion tokensCITE:E1. That combination — full parameter count for capacity, a much smaller active subset for efficiency, plus an open, commercially usable license — is how DeepSeek put frontier-class weights into public circulationCITE:E1.
How should the $5.576 million training cost figure be understood — what's included, what's excluded?
DeepSeek's own paper estimates the final training run at roughly $5.576 million, calculated using an H800 GPU rental rate of $2 per hourCITE:E2. The paper explicitly states this figure excludes prior research and the ablation experiments used to test architecture and data choicesCITE:E2. It is not a total R&D or capital expenditure figure. The widely repeated claim that "China built a frontier model for under $6 million" misreads this scope — the number covers one training run, not everything that produced the modelCITE:E2.
What's the reasoning performance gap between open-weight and top closed models?
DeepSeek-R1, released in January 2025, scored 79.8% on the AIME 2024 benchmark in its own paper, compared with OpenAI o1's reported 79.2% on the same benchmarkCITE:E3. That result is self-reported by DeepSeek and benchmarked specifically against o1, not against later frontier closed modelsCITE:E3. Third-party evaluation indices identify DeepSeek's models as the strongest currently available open-weight models, while still trailing the top closed models overallCITE:E3.
What did DeepSeek's market shock reveal about NVIDIA's valuation?
NVIDIA lost about $593 billion in market capitalization in a single trading day on January 27, 2025, as its shares fell about 17%CITE:E4. That single-day loss stands as the largest in US stock market historyCITE:E4. The sell-off reflected a market reassessment that a lower-cost, openly released model could reduce expected demand for high-end AI computeCITE:E4.
Why does chip hardware remain a generation behind — performance, power, and HBM/process bottlenecks?
Huawei's Ascend 910C delivers roughly 60% of NVIDIA H100's effective per-chip performance, according to third-party analysis, putting it about one generation behindCITE:E5. Huawei compensates at the system level: its CloudMatrix 384 architecture pools chips to close the gap in aggregate compute, but at a power cost of about 4.1 times NVIDIA's GB200 NVL72 systemCITE:E5. Upstream, the binding constraint is memory rather than logic capacity — Ascend chip production could theoretically reach about 800,000 units in 2025, but HBM supply limits push actual shipments down to roughly 200,000–300,000CITE:E6. China's leading foundry, SMIC, remains capped at a 7nm process node achieved via deep ultraviolet (DUV) multi-patterning without extreme ultraviolet (EUV) lithography, with yields still below industry normsCITE:E6.
Key figures at a glance
| Metric | Figure | Source |
|---|
| DeepSeek-V3 total / activated parameters | 671B total / 37B activated, 14.8T training tokens | CITE:E1 |
| DeepSeek-V3 reported training-run cost | US$5.576M (H800 at $2/hour; final run only) | CITE:E2 |
| DeepSeek-R1 vs OpenAI o1 (AIME 2024) | 79.8% vs 79.2% | CITE:E3 |
| NVIDIA single-day market cap loss (Jan 27, 2025) | ~US$593B, shares down ~17% | CITE:E4 |
| Ascend 910C per-chip performance vs H100 | ~60% | CITE:E5 |
| CloudMatrix 384 power draw vs GB200 NVL72 | ~4.1x | CITE:E5 |
| Ascend 2025 capacity vs. actual shipments | ~800,000 capacity vs. ~200,000–300,000 shipped | CITE:E6 |
| SMIC leading node | 7nm via DUV multi-patterning, no EUV | CITE:E6 |
| H20 license requirement | Effective April 9, 2025 | CITE:E7 |
| AI Diffusion Rule rescinded / Ascend warning | May 13, 2025 | CITE:E7 |
| H200 case-by-case review policy | Effective January 13, 2026 | CITE:E7 |
Taken together, the evidence traces a bifurcated position. On model architecture and open release, DeepSeek reached benchmark parity with OpenAI's o1 on AIME 2024CITE:E3 and preceded a record single-day market reaction in NVIDIA sharesCITE:E4, using a training-run cost of $5.576 million that its own paper says excludes total R&DCITE:E2. Yet the compute layer underneath has not closed the gap: Ascend 910C trails H100 by roughly 40 percentage points in per-chip performance while Huawei's system-level workaround draws about 4.1 times the power of NVIDIA's comparable rackCITE:E5, and actual Ascend shipments run at roughly one-quarter to three-eighths of estimated 2025 capacity because of HBM supply limitsCITE:E6. That hardware gap persists even as US export rules kept tightening through January 2026CITE:E7.