Sakura Internet May Raise Investment Up to Sevenfold on AI Data Center Demand

According to Bloomberg, Kunihiro Tanaka, CEO of Osaka-based data center provider Sakura Internet, said the company may increase this fiscal year's capital spending to roughly seven times its original plan to meet AI demand. Specifically, the company is considering expanding its capital investment plan from the initially budgeted 4.4 billion yen to somewhere between 20 and 30 billion yen.

Demand for AI infrastructure in Japan is growing rapidly, and expanding data center capacity has become an urgent priority. Sakura Internet has long offered GPU cloud services, but demand appears to be far outpacing its initial projections.

Huawei Announces Semiconductor Manufacturing Breakthrough, Narrowing the Gap with TSMC

A separate Bloomberg report says Huawei has announced a breakthrough in chip manufacturing technology and is narrowing the technology gap with TSMC.

While the detailed technical specifics are hard to glean from the coverage, Huawei is racing to secure its own supply of advanced semiconductors under US export controls. As global demand for AI chips continues to expand, Huawei's progress in semiconductor technology deserves attention, including its geopolitical implications.

Lens: An Open-Source Image Generation Model Matching Competitors at 3.8B Parameters

A 3.8B-parameter text-to-image foundation model called Lens has been published on Hugging Face Daily Papers. Its biggest strength is that it achieves quality on par with larger competing models while drastically reducing training compute.

Specifically, it was trained on 128 A100 GPUs and reaches competitive quality at roughly 19.3% of the compute cost of Z-Image. Three variants are available — Lens-Base, Lens-RL, and Lens-Turbo — with Lens-Turbo generating 1024-resolution images in 0.84 seconds. It supports resolutions up to 1440 and flexibly handles aspect ratios from 1:2 to 2:1.

It is released as fully open source, and the entire training pipeline is transparently disclosed — another point in its favor.

StepAudio 2.5: A Foundation Model Unifying Speech and Language Arrives

Also on Hugging Face Daily Papers, the technical report for StepAudio 2.5 has been released. StepAudio 2.5 integrates three capabilities — speech recognition, speech generation, and speech understanding — into a single foundation model, reportedly matching or exceeding the performance of dedicated systems for each task.

Multimodal models that handle speech and text in a unified way are expected to find wide-ranging applications, from real-time translation and voice assistants to content creation. Detailed benchmarks require consulting the paper, but its performance as a unified model is worth watching.

SkillOpt: The First Method to Systematically Optimize LLM Agent Skills

SkillOpt, published on Hugging Face Daily Papers, is a text-space optimizer that systematically optimizes the "skills" LLM agents use. Previously, agent skills were either crafted by hand or simply self-revised once generated. SkillOpt controls edit operations on skill documents with a "learning rate" and applies changes only when an improvement has been verified.

Across 52 evaluation cells (combinations of models, benchmarks, and execution environments), it achieved the best or tied-best result in every single one. On GPT-5.5, it delivered an average improvement of +23.5 points over the no-skill baseline, +24.8 points within Codex, and +19.1 points within Claude Code. The optimized skills were also confirmed to transfer across models and execution environments.

---