89% of AI Startup Revenue Goes to Just Two Companies — Anthropic and OpenAI Dominate the Market

The numbers revealed by The Information's analysis are striking. While total revenue across AI startups has finally reached $80 billion, a full 89% of it is concentrated in just two companies: Anthropic and OpenAI.

What these figures show is that the "winner-takes-all" dynamic in the AI industry is even more extreme than assumed. Despite the existence of hundreds of AI startups, a structure in which real revenue depends on two frontier companies raises questions about the sustainability of the industry as a whole.

Companies other than Anthropic and OpenAI are falling far behind on monetization. The reality that has come into focus is that while many startups have succeeded in raising funds, only a tiny fraction have managed to turn that into actual revenue.

The "All-You-Can-Eat" Era of AI Subscriptions Is Coming to an End

The "all-you-can-eat" model of flat-rate monthly plans for AI services is starting to crumble as price subsidies come to an end.

Until now, providers have heavily subsidized inference costs to win users, offering nearly unlimited usage for around $20 a month. But as usage grows, this model has become increasingly hard to sustain. A report from The State of Brand argues that the "time bomb" of these price subsidies has finally gone off.

For users, the flat-rate usage they've taken for granted will likely be gradually restricted, with a shift toward usage-based or tiered pricing. We've reached a turning point where everyday AI use is becoming something that requires cost awareness.

Unexpected AI Bills Pile Up on AWS and Google Cloud

According to The Register, a growing number of AWS and Google Cloud users are being hit with AI-related bills that far exceed their expectations.

Cloud providers' AI services offer poor visibility into usage, and developers frequently end up making expensive API calls without intending to. Agentic AI applications in particular suffer from a structural problem: autonomous, repeated calls cause unanticipated costs to pile up.

This issue is a textbook example of the "hidden costs" that come with serious production use of AI. Developers will need to put cost monitoring mechanisms in place as a matter of urgency.

Luce DFlash+PFlash Achieves 2.24x Speedup Over llama.cpp on an AMD 7900XTX

On Reddit's LocalLLaMA community, a reproduction report on a PR combining Luce's DFlash (flash attention for inference) and PFlash (flash attention for prefill) has been drawing attention.

On an AMD Radeon RX 7900 XTX (24GB VRAM), the user confirmed dramatic speedups on the Qwen3.6-27B model: **2.24x** decode speed and **3.05x** prefill speed compared to llama.cpp HIP.

The work used Lucebox's PR #119, showing that AMD GPU optimization on ROCm/HIP is making real progress. Not just for NVIDIA users but for AMD GPU users as well, local LLMs are becoming significantly more practical.

MoE Models' Refusal Layer May Be Hiding "Dialect-Conditional Safety Failures"

An interesting research finding was posted on r/MachineLearning. In experiments, MoE (Mixture of Experts) models behaved very differently toward prompts written in AAVE (African American Vernacular English) versus semantically equivalent prompts in academic English.

Using Qwen3.5-35B-A3B and a derivative model with its refusal layer removed, the researchers made three key findings. While the refusal layer is active, the model appears safe on the surface — but once it is removed, a marked degradation in safety on AAVE prompts becomes evident.

In other words, the MoE model's "refusal layer" may not be ensuring safety at all; it may simply be papering over dialect-dependent differences in behavior. The study is drawing attention as research that raises fundamental questions about how AI fairness and safety are evaluated.

CiteVQA: A New Benchmark That Exposes "Evidence Hallucination" in Document AI

CiteVQA, featured on Hugging Face's Daily Papers, sheds light on an overlooked problem in multimodal LLMs.

Current evaluation methods measure only the accuracy of the final answer, which misses a critical flaw: models frequently produce correct answers while **citing completely wrong locations as their evidence**.

In high-stakes domains such as healthcare, law, and finance, producing the right answer is not enough — knowing "where that answer is grounded" is essential. CiteVQA is a benchmark of 1,897 composite questions that evaluates both answer accuracy and evidence faithfulness, and it provides a mechanism for human visual verification through bounding-box-level citations.