NVIDIA Opens Joint AI Research Lab with South Korea's KAIST — National-Scale AI Strategy Gains Momentum

At this week's AI Summit in San Francisco, NVIDIA and South Korea's KAIST (Korea Advanced Institute of Science and Technology) announced the establishment of a joint AI research lab. The lab, to be housed at the Kim Jae-chang AI Graduate School at KAIST, will be the first joint AI research center between a Korean university and a global technology company.

The summit — attended by South Korean President Lee Jae-myung along with leaders of major companies and researchers — follows Jensen Huang's visit to South Korea last month. The lab aims to advance agentic AI by combining NVIDIA's full-stack AI expertise with KAIST's academic achievements.

In addition, a framework will be established to give major Korean companies SK Telecom, KT, and LG U+ access to NVIDIA's validation infrastructure. South Korea is pursuing a national strategy across every layer — AI infrastructure, specialized talent, and industrial applications — and its ambition to become a global AI innovation hub is unmistakable.

Microsoft Switches GitHub Copilot and Excel to Its Own "MAI" AI Models

Microsoft has begun full-scale deployment of its in-house "MAI" AI model series in GitHub Copilot and Excel. While continuing to use frontier models from OpenAI and Anthropic alongside them, the company will route high-frequency tasks to its lower-cost in-house models to improve cost performance.

The move is a telling example of big tech starting to rethink the cost structure of AI usage. AI features have historically depended entirely on state-of-the-art models, but a "routing" approach — matching the model to the nature of each task — is becoming the mainstream. Microsoft has indicated that it plans to extend this approach to other services going forward.

Copilot Studio: LLM Choice Can Mean a 100x Price Difference — Watch Out at Design Time

In Copilot Studio, which lets you build AI agents within Microsoft's Power Platform, an issue has been drawing attention in the community: LLM tier selection can drive costs up to 100x.

With the Basic tier priced at 1x, the Standard tier costs 15x and the Premium tier 100x. Even for an AI agent handling simple inquiry responses, specifying a Premium-tier model (such as GPT-4o) can result in costs far exceeding expectations.

In production, choosing the right tier for the task's complexity is essential. Particularly in scenarios where large volumes of requests are expected, Basic or Standard is often sufficient — and model selection at the design stage translates directly into cost.

"ai&" Launches an AI Inference Platform for Japanese Enterprises

ai& has launched "ai& Inference," an AI inference platform designed for Japanese companies. Positioned as an alternative to AI from overseas vendors such as Claude and GPT, the company claims it can dramatically cut AI usage costs.

For Japanese companies with strict data sovereignty and compliance requirements, a domestic vendor's inference platform can be a practical option. That said, its performance compared with overseas frontier models should be carefully verified against actual use cases.

What Is GEO, the New Optimization Approach for the AI Search Era?

As AI-powered search and answer generation (Generative Engines) become widespread, attention is growing around "GEO" (Generative Engine Optimization), an optimization approach distinct from traditional SEO.

GEO is the practice of ensuring that AI search engines correctly reference and cite your site. Whereas traditional SEO centers on keywords and links, GEO emphasizes structured content, clarity of information, and citability. Case studies involving real companies have already been reported, and the practice is entering practical use.

As AI search's share of traffic continues to expand, understanding GEO as a foundational web marketing strategy will only grow in importance.

Other Notable Developments

Debate over small MoE experts**: A discussion is heating up in Reddit's LocalLLaMA community: "If MoE's small experts (3B/4B/9B) work, why not just build small specialized models instead?" The fundamental question of whether large general-purpose models or small specialized models are more efficient has been reignited.

Meiji Yasuda Life's AI-powered effort estimation**: The insurer ran a PoC using AI agents to streamline system development effort estimation, a task that tends to rely on a small pool of experts. It is drawing attention as a Japanese example of applying AI to standardize highly person-dependent processes.

Bulk generation of AI ad copy**: Methods for bulk-generating creative copy with AI for Google and Meta ads are becoming practical, with cases of CVR improvements already reported.

Summary

Today's developments highlight two trends: "cost optimization" in AI usage and the "deepening of practical adoption." Microsoft's move to in-house models and the Copilot Studio pricing issue show that AI costs have become a business concern that can no longer be ignored. At the same time, efforts to embed AI into concrete business processes — such as GEO and workflow automation with AI agents — are accelerating. Combined with South Korea's national strategy, the competition over industrial AI applications is set to intensify further.