美银Arya研报称开放权重模型运行成本更高、利好半导体，但这是无个股增量事实的主题观点

Nvidia ( NVDA ), Intel ( INTC ), Marvell Technology ( MRVL ), Micron ( MU ), Lam Research ( LRCX ), and Synopsys ( SNPS ) are among the semiconductor companies that are poised to benefit as open-weight models turn out to be pricier to run than initially thought, Bank of America said. “We disagree with the sentiment that the shift to cheaper open-weight models, many of them Chinese, caps the AI buildout,” Bank of America analyst Vivek Arya wrote in a note to clients. “On OpenRouter, open-weight models have grown from ~27% of tokens in CY25 to ~71% CY26 YTD (~69% excl. free-tier usage). Over the same period, cloud capex is set to nearly double, from ~$510bn to ~$1.0Tn in CY26E (+98% YoY) and grow another +37% in CY27E.” Arya added there are three reasons for this: open models add to token demand and don't replace closed ones; open is how the weight is distributed and has no impact on the cost to run the models; and capex is funded by contracted backlog and usage-based pricing. “We acknowledge OpenRouter skews to developer and price-sensitive traffic, but point to company disclosures corroborating this trend, with Google's monthly tokens up ~7x YoY to [more than] 3.2 quadrillion,” Arya added. “In our view, open weight shifts margin away from model vendors, not demand away from silicon: the risk sits with model/API economics, not semis.” Additionally, Arya said the frontier labs, such as OpenAI ( OPENAI ) and Anthropic ( ANTHRO ), are still spending heavily on tokens, as the average weekly tokens on OpenRouter rose 93 times from the first quarter of 2025 to the third quarter of 2026. “As open models narrow the gap to the frontier, we believe closed labs must defend their lead via larger training runs, more RL/synthetic data and heavier test-time reasoning, i.e., the cost of staying ahead rises as the gap narrows,” Arya added.