The Unbundling of AI Intelligence: From General-Purpose LLMs to Function-Specific Optimized Architectures

Over the past three years, the core of LLMs such as GPT, Claude, and Gemini was Bundling, which integrated diverse functions like document extraction, search, and reasoning into a single general-purpose product. However, as it became clear that the latency and cost requirements for each function are fundamentally different, a Great Unbundling is underway toward architectures that offer economic rationality.

Changes in Cost Structure and Architecture

As agents convert a single human goal into hundreds to thousands of mechanical decisions, the cost differences at each stage compound into the overall product economics. Consequently, the architecture is shifting from 'frontier by default' to 'cheap by default, frontier on exception'. Search, reranking, memory, and computer use are being separated into distinct infrastructure layers, while small open-weight models absorb routine tasks.

The 'Decoder Tax' and the Rise of Jev

Inefficiencies of the 'Decoder tax' are being highlighted, where even simple classification or verification tasks require generating sentences via prompts before converting them back into decisions. Jev, launched by TypeSafe, isolates these judgment functions, returning structured decisions and probabilities instead of open-ended text. This is an example of packaging a single cognitive task with the cost and speed suitable for an agent's inner loop.

Adverse Selection in Frontier Models and the Rise of the Application Layer

As routing optimization removes cheap, predictable, high-frequency tasks such as iterative control and general generation from frontier models, these models increasingly receive only ambiguous and long-term (high-order) problems. This induces Adverse selection, which may reduce call frequency but could increase the value per call and the required test-time compute.

New Possibilities Created by Low-Cost Inference

When the cost of judgment drops to a fraction of a cent, it becomes possible to evaluate 100% of all support conversations, transactions, and contract clauses. The application layer acts as an 'Intelligence compiler', decomposing human goals into cognitive operations, purchasing the cheapest sufficient intelligence for each step, and recombining them. Intelligence is transforming from a specific event into a background property of software, with the next decade expected to focus on decoupling and recombination at the application layer.