⏱️ Lectura: 9 min
In April 2025, Nvidia declared a $5.5 billion accounting loss on a single product: the H20, the GPU designed to keep selling AI chips in China within Washington’s limits, which ended up banned too. That figure sums up the tension that’s still alive in 2026.
📑 En este artículo
- TL;DR
- What happened
- Context and History of Scientific AI in China
- Technical Details and Performance
- How to Verify It Yourself
- Impact and Analysis
- What’s Next
- Frequently Asked Questions
- What chip is replacing the Nvidia H20 in Chinese labs?
- Why did the US ban the H20 if it already complied with the 2023 rules?
- How much did it cost to train DeepSeek-V3 according to its own paper?
- Can SMIC manufacture chips at 5 nanometers or smaller without EUV machines?
- What is the Chinese government’s “AI+” initiative?
- Are Chinese models like DeepSeek available outside China?
- References
Despite the ban, scientific AI in China keeps advancing: labs and universities use models every day for research, with fewer Nvidia GPUs and more homegrown chips, like the Huawei Ascend, SMIC processes without EUV lithography, and models trained with efficiency tricks like DeepSeek’s.
TL;DR
- Chinese labs use agentic AI to accelerate scientific research despite US restrictions on advanced chips (C&EN, July 2026).
- In April 2025, the US expanded restrictions on the Nvidia H20, the chip designed to comply with the 2023 export rules.
- The Huawei Ascend 910B and 910C, manufactured by SMIC without EUV lithography, are the foundation of China’s domestic training clusters.
- DeepSeek-V3 reported a final training cost of about $5.576 million using H800 GPUs, despite the restrictions.
- SMIC has been producing the Kirin 9000s at 7nm with DUV multipatterning since 2023, without access to ASML’s EUV machines.
- Huawei’s CloudMatrix 384 cluster connects hundreds of Ascend 910C chips to approximate the aggregate performance of Nvidia GB200 systems.
- In 2025, China unveiled the state initiative “AI+” to integrate AI models into science, industry, and healthcare.
What happened
A report from Chemical & Engineering News describes how Chinese labs are deploying agentic AI systems on domestic supercomputing to speed up tasks like materials design, protein folding, and experimental data analysis, even with limited access to Nvidia’s most advanced GPUs.
The strategy combines two fronts. On one hand, universities and centers like the Chinese Academy of Sciences are adopting language models as research assistants, capable of generating hypotheses, reviewing literature, and automating simulations. On the other, China’s semiconductor industry is trying to close the hardware gap left by US export controls.
This approach isn’t exclusive to academia. State-owned energy companies and local pharmaceutical firms are also testing AI assistants to speed up trials and optimize industrial processes, replicating in China the same pattern of agentic adoption already seen in US and European labs, just on a different hardware stack.
Context and History of Scientific AI in China
The controls started in October 2022, when the US Department of Commerce banned selling China GPUs above a certain performance and interconnect threshold. Nvidia responded with cut-down versions, the A800 and the H800, designed to fall just below the legal limit.
In October 2023, Washington tightened the rule and those versions were banned too. Nvidia then designed the H20, a chip with less interconnect bandwidth but memory similar to the H100, built exclusively to keep selling in China under the new framework. In April 2025, the US government expanded the criteria again and blocked the H20 as well, forcing Nvidia to recognize billions of dollars in charges for stranded inventory.
The domestic response isn’t new. Huawei launched the Ascend 910B in 2023, and SMIC manufactures it with a 7-nanometer process that relies on deep ultraviolet (DUV) lithography with multiple passes, without access to the EUV machines that the Netherlands and the US block for SMIC. That same process had already been used in 2023 for the Kirin 9000s in the Huawei Mate 60 Pro.
flowchart TD
A["US export controls (2022-2025)"] --> B["Nvidia limits H100, H800, and H20 for China"]
B --> C["Chinese labs and universities"]
C --> D["Domestic chips: Huawei Ascend + SMIC"]
C --> E["Efficient models: DeepSeek"]
D --> F["CloudMatrix clusters"]
E --> F
F --> G["AI applied to science (materials, proteins, data)"]Technical Details and Performance
The central limitation isn’t just how many operations per second a chip does, but how much bandwidth it has to talk to thousands of neighboring chips during distributed training. That’s where the H20 and the Ascend lose out to the H100 or the B200. These numbers matter because they define how fast scientific AI in China can scale without depending on Nvidia.
| Chip | Manufacturer | Process | Key Limitation | Typical Use |
|---|---|---|---|---|
| Nvidia H100 | TSMC (4N) | ~4 nm | Not exportable to China since 2023 | Frontier training outside China |
| Nvidia H20 | TSMC (4N) | ~4 nm | Banned in April 2025 | Inference and moderate training (historical) |
| Huawei Ascend 910C | SMIC | 7 nm (DUV multipatterning) | Lower yield per wafer and higher manufacturing cost | Domestic training and inference in China |
⚠️ Heads up: the bottleneck isn’t just the individual chip, it’s the interconnect: without a high-bandwidth NVLink equivalent, scaling a cluster of thousands of accelerators requires more chips and more networking to match the same aggregate performance.
Huawei compensates for that gap with scale: its CloudMatrix 384 cluster connects hundreds of Ascend 910C chips with a proprietary interconnect mesh to approximate, in aggregate performance, systems like the Nvidia GB200 NVL72, though with higher power consumption per unit of useful compute.
On the software side, DeepSeek showed in 2024 that a frontier-class model can be trained while spending less compute than expected. The DeepSeek-V3 technical report reports a final training cost of approximately $5.576 million using H800 GPU clusters, thanks to techniques like FP8 quantization and a more efficient mixture-of-experts router.
📌 Note: that DeepSeek figure covers only the final training run reported in the paper, not the total cost of research, staff, or prior experiments: several analysts consider it a lower bound, not the full real cost of the project.
How to Verify It Yourself
To find out which GPU a server has available and whether it’s a limited version for China, two commands are enough.
nvidia-smi --query-gpu=name,compute_cap,memory.total --format=csvThis prints the exact chip name (for example NVIDIA H20 or NVIDIA H100 80GB HBM3), its CUDA compute capability, and total memory: the three data points that determine whether a chip falls into the categories regulated by the US Department of Commerce.
To track which Chinese institutions are publishing research on AI applied to science, you can query arXiv’s public API directly:
import urllib.request
url = (
"http://export.arxiv.org/api/query?"
"search_query=abs:agentic+AND+abs:science+AND+abs:China"
"&sortBy=submittedDate&sortOrder=descending&max_results=10"
)
respuesta = urllib.request.urlopen(url).read().decode("utf-8")
print(respuesta[:500])The script returns the last ten abstracts combining agentic AI, science, and Chinese affiliation, useful for verifying firsthand the volume of publications the report describes.
Impact and Analysis
The result isn’t that China matches Nvidia chip for chip: it’s that it stops needing to in order to advance applied science. A model trained efficiently on domestic hardware, even if it takes longer or consumes more energy per useful unit, still works to speed up designing a catalyst or analyzing a genomic dataset.
That distinction matters for export policy: each new round of controls pushes China to invest more in its own semiconductor supply chain, with SMIC as the critical bottleneck because, unlike chip design, manufacturing at scale without EUV lithography is much harder to replicate quickly.
The lesson extends beyond China. Any country without access to Nvidia’s most advanced chips (India, some players in the Middle East, even parts of Europe with limited quotas) is watching closely whether software efficiency can compensate, at least partially, for a hardware disadvantage. If the answer is yes, export controls lose part of their deterrent effect in the medium term.
What’s Next
The open question is whether Huawei and SMIC can scale Ascend production faster than Washington adjusts its list of restricted chips. The race to scale scientific AI in China now depends as much on SMIC’s wafers as on Washington’s policy.
The state initiative “AI+”, unveiled in 2025 to integrate AI models into industry, healthcare, and science, depends directly on that supply race.
📖 Summary on Telegram: View summary
Try it yourself: run nvidia-smi --query-gpu=name,compute_cap --format=csv on your own cloud instance to see exactly which chip you’re being billed for.
Frequently Asked Questions
What chip is replacing the Nvidia H20 in Chinese labs?
Mainly the Huawei Ascend 910B and its successor, the 910C, manufactured by SMIC with a 7-nanometer process without EUV lithography.
Why did the US ban the H20 if it already complied with the 2023 rules?
Because in April 2025 the government expanded the restriction criteria, which led Nvidia to recognize billions of dollars in charges for stranded inventory.
How much did it cost to train DeepSeek-V3 according to its own paper?
The technical report reports a final training cost of approximately $5.576 million on H800 GPUs, not counting prior research or staff.
Can SMIC manufacture chips at 5 nanometers or smaller without EUV machines?
Not with competitive density or yield. It uses multipatterning with DUV lithography at 7 nanometers, a more expensive process with lower yield per wafer.
What is the Chinese government’s “AI+” initiative?
A state plan unveiled in 2025 to integrate artificial intelligence models into industry, science, healthcare, and public administration.
Are Chinese models like DeepSeek available outside China?
Yes, several are released with open weights under permissive licenses, allowing them to be run and audited outside Chinese infrastructure.
References
- Chemical & Engineering News: report on Chinese labs using agentic AI in science despite chip limitations.
- DeepSeek-V3 Technical Report (arXiv): technical paper with the training cost breakdown and model architecture.
- Wikipedia: SMIC: history and process capabilities of the Chinese semiconductor foundry.
- Nvidia Newsroom: official announcements on products and export control compliance.
📱 Like this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.
Imagen destacada: Foto de Michael Myers en Unsplash
0 Comments