NTT DATA to Quadruple Domestic Data Centers with $9 Billion to Meet AI Demand
NTT DATA Group is moving ahead with a plan to invest at least $9 billion (roughly 1.3 trillion yen) through 2033 to quadruple its computing capacity to 1 gigawatt. According to reports citing people familiar with the matter, the company — Japan's largest data center operator, part of the government-backed NTT — intends to add roughly 750 megawatts of capacity over the next seven years, positioning itself to capture surging demand from companies falling behind in AI development.
Behind the move is rising demand for AI computing power in Japan. Securing data center capacity to support everything from model training to inference has already become an international point of competition, and this investment lets a domestic operator respond directly to that infrastructure demand. While the reports cover the investment amount and capacity targets, details such as the locations of new facilities and construction schedules remain unspecified, so further announcements will be needed.
Inference Engineering, the Discipline That Comes After "Training" — The Front Lines According to Baseten
The field of inference engineering is rapidly maturing into a distinct specialty separate from model training. According to engineers at Baseten, a leading inference infrastructure company that raised a massive Series F round, the heart of the job is "turning the weights you got from training into products that are fast, cheap, and reliable." Optimization options are layered and combined: cache-aware routing that eliminates redundant computation by reusing KV caches, disaggregated setups that split prefill (which processes input) and decode (which generates output) across separate GPU pools, and speculative decoding that lets a small model run ahead to accelerate a large one.
One particularly emblematic effort is "grafting" different open-source components into a single model. A highlighted approach takes GLM-5.2, which lacks vision capabilities, attaches Kimi's vision encoder, and trains only the connecting layer (the projector) between the two. The design keeps the "LLM as the brain" and the "encoder as the eyes" frozen and learns only the part that bridges them. When trained not just to describe the contents of images but to answer questions about them, a phenomenon known as "grokking" — where understanding gradually deepens — was reportedly observed. Another cited advantage: without an image provided, the original GLM-5.2's performance remains unimpaired.
Quantization is also being reconsidered. In one experiment, quantizing more of the model actually improved throughput by about 20% while preserving benchmark quality, because errors introduced in different layers canceled each other out. Inference optimization is said to still leave room for 20% to 200% speedups — up to 10x when things go well. At the same time, the piece points to real-world difficulties: failures such as "mode collapse," where the model repeatedly outputs the same token, often stem from software or hardware rather than the weights themselves, and differences in interconnect speed between clusters can expose race conditions.
Salesforce Japan Runs 300 AI Agents Autonomously, with the "4R" Organizational Overhaul
While only 11% of companies have managed to deploy generative AI across the whole organization, Salesforce Japan is reportedly running some 300 AI agents autonomously in-house. The company says it aims to go beyond individual productivity gains and elevate the effort into company-wide transformation (AX), with the keys being clear boundaries that prevent unmanageable "stray agents," a "4R" framework for restructuring the organization and its people, and a management model that oversees both humans and AI.
The worry that individual users adopting agents in an uncoordinated way undermines governance is a persistent challenge when enterprises roll out AI in earnest. The idea of bringing agents into an organization-wide framework instead of letting them run loose could serve as a useful reference for Japanese companies seeking to make sure generative AI doesn't end as "isolated workflow improvements." That said, the article mainly conveys the direction of the company's efforts, and the parts delving into the specifics of the 4Rs or quantitative verification of the results are limited.
AI-Proctored Remote Exam Hits a Glitch, Forcing 58,000 Students to Retake
A serious malfunction struck an AI-proctored remote exam, and it has been reported that 58,000 students are being forced to retake the test. The affected exams were remote tests monitored by an AI-based proctoring system, but as for what actually went wrong, the reports' headlines indicate little beyond the fact that "there was a failure serious enough to require retakes."
The episode once again underscores how difficult it is to deploy AI proctoring in settings that directly affect exam fairness and test-taker burden. As AI adoption spreads in education, the risk that automated-monitoring misjudgments or technical glitches trigger large-scale do-overs is a point of debate that cannot be ignored. Further details on the cause and preventive measures may be reported in the original article.
NVIDIA Releases Full-Duplex Voice Conversation Model "VoiceChat-11B"
NVIDIA has released "NemotronLabs VoiceChat-11B," a model supporting full-duplex voice conversation, on Hugging Face. "Full duplex" refers to a communication method in which sending and receiving happen simultaneously — a property said to reduce conversational gaps and enable more natural responses in voice AI. Released at a scale of roughly 11 billion parameters, it is an open voice model that the community has begun trying out.
Highly real-time voice AI is an area that shapes the agent and assistant experience, and the arrival of a full-duplex model looks set to accelerate the race toward practical deployment. On the other hand, this release centers on the model card and listing information; details such as concrete performance figures, usage terms, and supported languages are not spelled out, so actual behavior will need to be verified separately.