AI Competition Shifts to Task Cost and Productivity, IITP Says

The Institute of Information & Communications Technology Planning & Evaluation (IITP) named "proving AI's potential through AX productivity" the central theme for the AI and ICT industry in 2027, releasing its "Top 10 AI and ICT Issues for 2027" outlook on the 11th.

The analysis holds that industrial competitiveness will hinge on whether AI can carry out complex tasks at the lowest possible cost. Competition in artificial intelligence is shifting from improving raw performance to delivering measurable productivity gains at actual industrial sites.

In plain terms: the winner is whoever completes a unit of work for the least money, not whoever posts the highest benchmark score.

What is the agentic economy?

The most notable change IITP identifies is the spread of what it calls the "agentic economy," in which AI moves beyond carrying out human instructions to completing tasks on its own. The institute expects AI to perform complex missions based on long-term memory and on collaboration among multiple agents, forming an economic ecosystem in which agents transact autonomously.

Autonomy brings new requirements. As AI gains independence, IITP said, systems that verify an agent's identity and authority, and that clearly assign responsibility for completed work, will grow in importance.

The takeaway: agents will trade with each other, and someone has to vouch for who they are and who pays for their mistakes.

What changes in physical AI and robotics?

IITP also selected the mainstreaming of physical AI — AI operating directly in the real world — as a major issue. Robots will move beyond repeating preset motions to performing complex tasks in environments they encounter for the first time and recovering from errors on their own, the institute said.

Technology that accumulates and reuses experience data collected from robot behavior and surrounding environments will become important to reaching that goal.

Could AI start improving itself?

Recursive self-improvement (RSI), in which AI advances AI itself, is expected to draw attention. IITP said advances in RSI could bring forward the realization of artificial general intelligence (AGI) and artificial superintelligence (ASI).

The institute also projected that AI will evolve into an "autonomous scientist" that sets hypotheses and conducts experiments and verification on its own, applied to humanity's hardest problems such as disease and food crises.

How will infrastructure competition change?

In AI infrastructure, competition over cost and efficiency is expected to intensify. IITP laid out two fronts:

• Chips: AI chips will move away from boosting the performance of a single chip toward combining different semiconductors, such as GPUs and neural processing units (NPUs), and optimizing software and systems together. For domestic NPUs in particular, the key task is securing competitiveness in low-power inference by jointly optimizing AI models and inference engines.
• Data centers: Competition to raise both construction speed and operating performance at AI data centers (AIDCs) will begin in earnest. That means shortening build times with modular data centers and running power and cooling systems efficiently to maximize AI token output per unit of electricity.

IITP flagged a specific economic risk: even if token prices for AI models fall, the total cost of carrying out complex work can rise. Cutting "task cost" — the expense of completing an individual mission — will emerge as a key competitive factor.

What else made the top 10 list?

IITP selected four additional issues for next year:

• The evolution of spatial computing centered on AI glasses
• "AI-native 6G," in which AI is embedded in network operations and services
• A "cyber shield dome" using autonomous security agents
• "Full-stack sovereign AI" that secures control capabilities across every layer from semiconductors to networks and security

The common thread across all ten issues: autonomy, verification and unit cost now define the agenda, and raw model performance no longer does.