Q.ANT's Photonic Computing GPU — From Transistors to Light

Optical (photonic) computing technology developed by German company Q.ANT is attracting attention in Reddit's LocalLLaMA community.

The company claims its photonic processors achieve **50x the compute performance** and **30x the power efficiency** compared to conventional transistor-based chips. They are currently in volume production operation at the Leibniz Supercomputing Centre in Munich, and the second generation is touted to deliver 100x performance and 90x power efficiency.

Some argue that the same shift that saw fiber optics replace ADSL is now underway in GPU architecture. The company has brought on former IBM executive Bruno Spruth as CTO and has opened a US base in Austin, Texas.

Note, however, that these performance figures are based on the company's own announcements, and no independent third-party verification has been confirmed at this point. Photonic computing has been a field of research for many years, and its progress toward commercialization will continue to warrant close attention.

KGC 2026 — Knowledge Graphs Emerge as Production Infrastructure Filling Vector DB Gaps

The Knowledge Graph Conference 2026 (KGC 2026), held in New York, showcased numerous production knowledge graph (KG) systems from enterprises.

According to a report shared on Reddit's MachineLearning community, the following implementations were presented:

Bloomberg**: Released formal dependency models for ontology governance

AbbVie**: Built "ARCH", an internal KG for drug and disease area intelligence. It connects a scoring engine, researcher dashboards, and an LLM companion, with the KG functioning as a trusted single source of truth

Morgan Stanley**: Automatically runs continuous SHACL drift detection on risk reporting data on a weekly basis

The reporter concluded: "Knowledge graphs are functioning as infrastructure, not a search layer. The graph does the reasoning work — it's not just lookups." They argue this is clear evidence of the limitations of relying solely on vector DBs.

All decks from the conference have been made publicly available on Google Drive.

SoftBank Group Posts Over ¥5 Trillion in Net Profit — Highest Ever for a Japanese Company

On May 13, SoftBank Group (SBG) announced that consolidated net profit for the fiscal year ended March 2026 reached **¥5,022.7 billion** (up 333.7% year over year).

Gains on its OpenAI investment contributed significantly, and the numbers reflect strategic investments in the AI space beginning to pay off.

Recursant — Mesh-Based AI Agent Governance

An open-source AI agent governance platform called "Recursant" was featured on Hacker News.

It applies Istio's sidecar pattern to provide a mechanism for isolating and controlling agents at the network layer. For large enterprises that need to manage AI agents running across different frameworks and runtime environments in a unified way, it is expected to reduce compliance risks.

As an approach to achieving agent governance while avoiding vendor lock-in, the project is currently at the stage of soliciting feedback from the community.

Attempting Agent Coding on 8GB VRAM with BeeLlama.cpp

Efforts to run agent coding on 8GB VRAM GPUs using the llama.cpp fork "BeeLlama.cpp" are gaining momentum in the LocalLLaMA community.

Target models include Qwen3.6-35B-A3B, Qwen3.6-27B, Gemma-4-31B, and Gemma-4-26B-A4B, leveraging features such as DFlash (inference acceleration) and TurboQuant. There are also reports of running Qwen 3.6 27B Q5 with a 200k context on an RTX 3090.

Tokens-per-second measurements are being shared as the community explores just how far practical agent coding can go on an 8GB VRAM setup.