SoftBank Hits All-Time High — Surges 40% in 5 Days on OpenAI IPO Expectations
SoftBank Group's stock rose 4.6% in Tokyo trading on May 25, setting a new all-time high. In just five days since May 20, the stock has climbed 40%, pushing the company's market capitalization above ¥40 trillion (approximately $252 billion).
Behind the surge are expectations for IPOs of OpenAI and SB Energy. SoftBank is known as a major investor in OpenAI, and the market is heavily pricing in the potential returns if the company goes public. As optimism over AI demand lifts the entire tech sector, SoftBank has become a symbolic stock attracting capital.
However, neither the timing nor the terms of the listings have been finalized, and the rally is largely expectation-driven at this point. Whether actual investment returns can live up to market expectations will depend on how OpenAI's business unfolds.
IBM Announces Product Suite for Governing Thousands of AI Agents
As enterprises increasingly adopt AI agents, how to govern multiple agents has become a major challenge. To address this, IBM has announced an AI operating model along with a related suite of products.
The system manages and monitors agents at the scale of thousands, and is designed to control agent behavior from governance and security perspectives as well. The product lineup appears to be built with the "governance wall" enterprises face when seriously embedding AI agents into their operations in mind.
ServiceNow Announces Features to Fill the "Data Gap" for AI Agents
Like IBM, ServiceNow also announced new AI agent-related features at its annual event. What stands out is a set of capabilities that directly tackle the barriers to AI agent adoption: siloed data and immature governance.
For AI agents to function effectively in business, they need access to data scattered across various internal systems. In reality, however, data is fragmented between departments, creating a "data gap" that prevents agents from making sound decisions. ServiceNow's new features aim to bridge this gap and make agents more practical.
Qwen 3.6 Records 3500tps on RTX PRO 6000
Benchmark results for Qwen 3.6 are making waves in the local LLM community. In a test using two RTX PRO 6000 GPUs, Qwen 3.6 35B BF16 recorded a generation speed of 3500tps (tokens per second) at 128 concurrency. Prompt processing speed also reached 30,000tps.
The 27B model achieved 1800tps at 64 concurrency with MTP (Multi-Token Prediction) enabled, and recorded 1600tps even with MTP off — demonstrating how much the RTX PRO 6000's memory bandwidth contributes to fast inference with large models.
llama.cpp Checkpoint Fix Improves Agent Coding
A patch modifying checkpoint generation has been proposed for llama.cpp, aimed at improving responsiveness during agent coding.
Currently, tools such as opencode rewrite conversation history for context optimization, which forces llama.cpp to reprocess the entire context. With 70k tokens, this triggers "forcing full prompt re-processing" and causes long wait times.
The fix avoids the worst case (full prompt re-processing) and moves behavior closer to the best case, where only the parts that actually changed are reprocessed. The proposer has been using it for about two weeks and reports a clear improvement in the perceived responsiveness of agent coding.