Unverified50% confidenceTradeoffExact time
Liang Wenfeng estimates that training a top-tier large model on Huawei Ascend requires approximately 4 times the number of cards needed with NVIDIA.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/23/2026
First Seen
Valid until: 12/22/2026
Sources
Related Entities
Related Claims
VerifiedIndustry estimates suggest that the inference cost of mainstream frontier models is approximately several times to tens of times the training cost.66% similarUnverifiedBefore the era of large models, the largest model the presenter had trained had only approximately 4,000 parameters.65% similarUnverifiedA single AI inference request consumes little electricity, but because requests number in the billions, the cumulative energy consumption may exceed that of training itself.64% similarUnverifiedChinchilla, trained with 70 billion parameters and 1.4 trillion tokens of training data, surpassed Gopher—a model with four times its parameter count—on multiple benchmarks.63% similarUnverifiedThe NVIDIA H200 improves inference efficiency by approximately 3x compared to the A100 on specific tasks, but the simultaneous expansion of model parameter scale has almost offset the cost dividend brought by hardware progress.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/943522API
curl https://kongchang.com/api/v1/knowledge/claims/943522MCP
get_claim(id=943522)