Sources: Tencent, Alibaba, ByteDance, and other Chinese companies are ramping up orders for Nvidia's H20 AI chip due to booming demand for DeepSeek's models
Chinese companies are ramping up orders for Nvidia's (NVDA.O) H20 artificial intelligence chip due to booming demand …
Reuters
Context & Ripple Effects
Nvidia had already adapted its China offering with the modified H800 for Chinese customers, before opening orders for the China-specific H20 in 2024. The new buying wave shows that this product line is becoming tied to demand for deployed AI models, not merely chip availability.
Tencent, Alibaba, and ByteDance are the common buyers across that coverage, making their H20 orders a useful indicator of how model demand is translating into infrastructure procurement among China’s largest platforms.
First-order effects
Nvidia gains an immediate increase in H20 demand from major Chinese technology companies as they provision capacity for DeepSeek-model usage.
Tencent, Alibaba, ByteDance, and peers commit more spending and deployment planning to Nvidia’s China-specific AI hardware.
Second-order effects
The order surge reinforces the value of export-tailored accelerators, pressuring Nvidia to keep a viable China product roadmap while Chinese buyers weigh available alternatives.
Large-platform demand concentrates AI-compute purchasing among a small set of major buyers, potentially shaping which model providers and cloud services can scale fastest.
Third-order effects
If application and model adoption repeatedly drives hardware buying, AI infrastructure demand will be governed increasingly by inference and deployment workloads rather than solely by frontier-model training cycles.
The episode fits a longer-running split in which regional policy constraints encourage distinct hardware configurations and procurement strategies, even as Nvidia remains important to Chinese buyers.
The trend: This is one data point in the shift from generalized AI-chip investment to model-led infrastructure purchasing through region-specific hardware supply chains.
Deepseek found a non documented instruction set in Nvidia chips, found out it should be unreliable, but verified it was always reliable... and used it. I can just imagine them looking at the chip and zapping it with various instructions to see what each did. [image]
Introducing DeepSeek-R1 optimizations for Blackwell, delivering 25x more revenue at 20x lower cost per token, compared with NVIDIA H100 just four weeks ago. Fueled by TensorRT DeepSeek optimizations for our Blackwell architecture, including FP4 performance with state-of-the-art […
“Alleviate concerns”? Ya think? You can't make this stuff up. “The surge in orders, which is being reported for the first time by Reuters, underlines Nvidia's dominance of the market and could help alleviate concerns that DeepSeek might cause a slide in AI chip demand.”
@teortaxesTex I wouldn't trust the numbers from FT. It's all rumors and the reporters aren't chip experts. But directionally it's plausible that SMIC/HW are improving Ascend yields a lot. I think SemiAnalysis has even larger projections for 910C shipments in 2025.
FT on Huawei / SMIC Ascend 910C chip yields: doubled to 40% and now targeting 60%. TSMC H100 yield is ~60% ? When I was in Shanghai I had dinner with one of the top SMIC guys involved with leading edge chip fabrication. The team has a big Taiwanese component with experience at [i…
> Ascend 910C achieving a 40% yield, up from 20% a year ago. > plans to further improve yields to 60% > plans to produce 100K Ascend 910C, 300K 910B chips this year That's like 12K wafers, 1000 WPM. They're claimed to have 30K WPM for N7. Numbers don't work out. @dylan522p