"Tokenpocalypse" — Will AI Companies Going Public Drive Up Token Prices?

Anthony Ha at TechCrunch, in a notable article titled "Is this the dawn of the Tokenpocalypse?", flags the risk of rising token prices as major AI companies prepare to go public.

As major AI companies plan initial public offerings (IPOs), they need to demonstrate profitability to investors — and as a result, a wave of API price hikes could follow. The piece suggests that LLM costs, which have been falling rapidly so far, could reverse course under the pressure of going public.

On Japan's Qiita, an explainer on runaway AI costs has also been published. It argues for a shift from the era of "go fast" to "control with guardrails," and reports that a new initiative called the Tokenomics Foundation has launched. Optimizing AI costs has become a pressing issue for developers and enterprises alike.

Banking Industry Prepares for Mass Layoffs Tied to AI Adoption

According to Bloomberg, the banking industry is laying the groundwork for large-scale job cuts as AI is adopted in earnest.

The scale and timing of the cuts vary from bank to bank, but as AI-driven automation advances, significant restructuring of human resources is reportedly under consideration, centered on back-office and analytics roles. The financial industry has long stood at the forefront of digitization, but the rise of generative AI is beginning to reshape employment itself.

The Guardian Charts the Staggering Scale of AI Investment in Six Charts

In a special feature titled "Billions spent and hypothetical returns: the AI boom explained with six charts," The Guardian visualized the reality of the AI boom through data.

It lays bare the fact that while enormous sums are being invested, most of the returns remain hypothetical at this stage. It is a useful analysis for understanding the economic dynamics surrounding the AI industry — infrastructure spending, the war for talent, and surging demand for compute resources.

Qwen 3.6 27B Ranks 18th on the DeepSWE Benchmark — The Reality of Local LLMs

Benchmark results for Qwen 3.6 27B on DeepSWE were shared in Reddit's LocalLLaMA community.

The model scored 2% (rounded up from 1.79%), placing 18th out of 20 models. That said, it reportedly outperformed Claude Haiku 4.5 and Minimax M2.7. The benchmark took 70 hours on a single RTX 6000 Pro Blackwell, averaging 32 minutes per task and generating an average of 44k tokens of output.

The person who ran the benchmark noted that "even Kimi-k2.6, the strongest open-source model that can run locally, lags far behind state-of-the-art models." It was a stark reminder that despite continued performance gains, the gap between local LLMs and the frontier remains wide.

Open-Source MoE/MoD Training Framework for PyTorch Released

On Reddit's LocalLLaMA, a PyTorch-based training framework supporting MoE (Mixture of Experts) and MoD (Mixture of Depths) has been released under the Apache 2.0 license.

Key features include:

Custom CUDA kernels for RMSNorm, RoPE, SwiGLU, and MoE routing (2–7x faster than PyTorch on T4)

Support for hybrid configurations of MoE and MoD with up to 64 experts

An orchestrator that monitors more than 20 metrics and automates learning rate tuning as well as adding and removing experts

Support for configurations ranging from 500K to 300B parameters

Metal shaders for Apple Silicon included

It has been benchmark-verified on T4 (Google Colab), and a free Colab demo is also available. As an effort to build self-reliant training infrastructure for large-scale models, it is drawing attention in the community.

How Manufacturers Are Actually Using AI Agents — Real Voices from Japan

The results of the "製造業のAIエージェント活用実態調査 2026" (2026 Survey on AI Agent Usage in Manufacturing), conducted by ITmedia's MONOist, have been published.

Manufacturing has long been regarded as one of the industries furthest along in AI adoption — but how widely are AI agents actually deployed on the ground, and what challenges exist? The results reportedly reveal the gap between expectations and reality, along with concrete hurdles to adoption.