Japan industrial electronics sector: Points to watch ahead of results announcements

Research evidence excerpt

Japan industrial electronics sector: Points to watch ahead of results announcements

Nomura | Japan industrial electronics sector 28 July 2026

Open-weight models still require computing resources

The 17 July release of Kimi K3, which performs almost as well as US Frontier AI models

but at lower cost (Figure1), has resulted in plenty of discussions about the following two

points: (1) how will US Frontier AI models monetize themselves, and will this hit capex at

hyperscalers, now that improved performance at Chinese open-weight models has turned

them into rivals?; and (2) will increases in the share of Chinese open-weight models also

lead to fluctuations in market share at hardware suppliers? With respect to (1), we still see

near-term uncertainty as it is difficult to estimate the demand of open-weight models for

computing resources. We therefore think that market sentiment is likely to fluctuate

depending on capex plans and comments at hyperscalers. With respect to (2), we stress

that the use of open-weight models will not alter the suppliers of computing resources.

Fig. 1: Performance and cost by AI model

Note: As of 28 July 2026.

Source: OpenRouter, Nomura

While it is possible for users to run open-weight models on-premise using their own

infrastructure, they can also easily access APIs provided by inference providers,

particularly in the case of models with large parameters. As Figure2 shows, the main

inference providers are companies such as Hugging Face, Fireworks, and Together AI,

and hyperscalers themselves provide inference services. Furthermore, while most

inference providers offer API access to multiple models, by using LLM aggregators, users

can select from numerous providers for the same model based on speed and price

considerations.