Unverified50% confidenceFactExact time
当AI集群规模扩展到数万张GPU时，整体效率可能暴跌到40%以下，超过一半的算力被消耗在数据传输和等待过程中
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
7/2/2026
First Seen
Valid until: 9/30/2026
Sources
Marvell：AI基础设施幕后的隐形巨头，四张技术王牌解析
bilibili中天评论6/6/2026
Related Claims
UnverifiedWhen scaling to tens of thousands of GPUs, overall compute efficiency can plummet below 40%, with more than half of compute power consumed by data transfer and waiting — a phenomenon known as the 'communication wall.'80% similarUnverifiedAI模型的边际成本随规模扩大而降低，用户越多，通过批处理和GPU利用率优化后每次推理的平均成本越低79% similarUnverified在完整的AI应用系统中，GPU推理通常只占总计算量的30%-40%，其余60%-70%的计算负载由CPU承担76% similarUnverified当Token使用量超过60%-70%时，AI的输出质量会显著下降，推理能力明显降低74% similarUnverifiedAI推理成本主要构成是A100/H100等高端GPU的折旧与电力，属于资本密集型刚性支出，难以随收入下滑而快速压缩73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/51578API
curl https://kongchang.com/api/v1/knowledge/claims/51578MCP
get_claim(id=51578)