On September 15, TypeSafe AI unveiled its decision-oriented model, Jev, which secured $40 million in seed funding. Rather than handling conversational or text generation tasks, Jev focuses on three categories of structured judgment work: classification, selection, and scoring. It delivers probabilistic outputs in place of free-form text, with inference latency spanning roughly 70 to 500 milliseconds and averaging around 100 milliseconds.

Regarding pricing, the input cost through the Vercel AI Gateway stands at approximately $0.04 per million tokens. Official testing indicates that within decision-centric workflows such as classification, routing, scoring, and determining next steps, Jev achieves up to roughly 194 times faster performance compared to traditional generative large language models, while cutting costs to about 1/445 of previous levels.

As AI technology evolves from "generation" toward "judgment," E Fund AI (03489) and E Fund Asia Semiconductor (03486) combine to form a "software-hardware" pairing that captures the trend of layered collaboration within AI systems.

The launch of Jev points to a substantive shift in the division of AI inference responsibilities. General-purpose large models remain responsible for understanding, reasoning, and generation, while Jev takes on high-frequency, repetitive, and rule-defined judgment tasks. The two operate in a layered, collaborative structure within the system.

For use cases requiring massive decision-making calls, such as agents, browser automation, and content moderation, reductions in cost and latency directly determine whether these applications can scale. This mirrors the Jevons paradox: once judgment costs fall significantly, automation scenarios previously deemed uneconomical become broadly unlocked.

For computing power and infrastructure, the proliferation of decision-oriented models signals a further rise in total inference call volumes, rather than substitution. Within this context, E Fund AI (03489) provides coverage of US computing power leaders alongside Chinese AI applications, benefiting from both the computing power and application segments, and serves as a core instrument for tapping into the layered collaboration trend of AI systems.

Meanwhile, E Fund Asia Semiconductor (03486) focuses on memory, foundry, advanced packaging, and equipment. Rising inference call volumes transmit directly into memory and wafer manufacturing momentum, making it well suited as a hardware-side complementary allocation.

It is worth noting that Jev was primarily trained on English-language data, so its accuracy in Chinese semantic judgment may be comparatively lower. Real business samples should be used for testing before official deployment. Complex calculations and multi-hop reasoning fall outside its suitable use cases, and high-risk operations should still retain human review alongside rule-based safeguards.

In the near term, this event carries a positive marginal impact on both AI application layers and computing power demand. However, the actual pace of adoption for any single model remains something to monitor closely.