[DigitalToday reporter Chi-gyu Hwang] "Over the past 3 years, LLMs handled document extraction, search, ranking, anomaly detection, tool selection, web browsing, answer writing, verification and high-level reasoning in one model. Now that bundling is starting to unravel."

Foundation Capital partner Jaya Gupta (자야 굽타) summarised the trend as the "Great Unbundling of Intelligence" in a recent post on social media platform X (Twitter) and presented it as a point to watch.

She said computing requirements vary widely by capability, and models that excel also differ. She said the economically rational structure changes as gaps widen in performance, latency and cost, underscoring the spread of trends accelerating the unbundling of intelligence.

She particularly highlighted agents. Agents break a single human-set goal into hundreds or thousands of mechanical judgements, and small differences in cost and latency at each step accumulate and determine product economics, she said.

Gupta said, "Reasoning has now become an actual cost of goods sold, and small models and open-weight models have become good enough to handle everyday tasks. Search, memory and computer-operation functions have now broken away from LLMs and are being sold as separate services. Evaluation frameworks have also improved, making it possible to judge whether a cheaper system is sufficient."

As frontier models improve, unbundling can become a safe card from the user’s perspective.

She said, "Frontier models at the very top can handle hard problems that cheap models cannot solve. Going forward, it will shift from using frontier models from the start and later cutting costs to using cheap models as the default and passing only blocked problems to frontier models."

Gupta also cited TypeSafe’s 'Jev' as an example. According to her, Jev can be seen as a model that extracts only the judgement function. If given a complex situation, a narrow question and a list of possible answers, Jev provides decisions and probabilities rather than sentences.

She said, "What software needs are decisions such as whether it is urgent, whether to continue or stop, and whether to allow or block. But right now language models first write sentences and then convert those sentences back into decisions. A significant portion of what agents repeat internally does not need to write sentences."

At that point, an uncomfortable question is unavoidable from the perspective of LLM providers’ revenue structures. Anthropic earns large profits by charging high prices even for easy tasks, but if players emerge that handle easy tasks cheaply, Anthropic’s profits are bound to fall.

Gupta said, "A judgement-only model like Jev can decide whether something is urgent at a cost of less than 1 cent, and rerankers cover ranking. If search is handled by search companies and ordinary writing by cheap models, AI apps do not need to fight Anthropic."

According to Gupta, this situation can be summarised as 'adverse selection' in frontier reasoning. As routing improves, cheap, predictable, high-volume work peels away, leaving only ambiguous, long-horizon problems with frontier models, she said.

She sees new products becoming possible if the cost of judgement falls to under 1 cent and judgement time drops to the level of hundreds of milliseconds.

She said, "A judgement model that inspects results can be attached every time an AI agent acts. Judgement functions can be built into areas that require fast responses, such as operating a web browser or conversing by voice." She added, "It will be possible to evaluate 100 percent rather than sampling 1 percent of consultations, searches, transactions, insurance claims and contract clauses. Continuous supervision and continuous quality control become possible." She stressed, "Over the past few years, the industry has considered how many functions to pack into a single model. The next 10 years could instead be a time to peel those functions off one by one and recombine them as needed in applications."