Today's News
Cohere via Interconnects by Nathan Lambert
Cohere released Command A+, a 218B-A25B mixture-of-experts model under Apache 2.0 license, marking a significant licensing shift for the sovereign AI player. The model unifies capabilities across the Command family—reasoning, multimodal, tool use, and 48-language support—into a single deployable system optimized for agentic workflows. Command A+ runs on a single B200 GPU with 4-bit quantization or two H100s, delivers up to 63% higher output tokens per second than Command A Reasoning, and achieves 85% on τ²-Bench Telecom (up from 37%) and 25% on Terminal-Bench Hard agentic coding (up from 3%). Cohere's year-long deployment of North, its enterprise AI workspace, drove the model's design.
Simultaneously, Zyphra and Poolside committed to open-weight model releases as their default strategy, with both companies releasing flagship models under Apache 2.0: Zyphra released 74B-A4B and 8B-A0.6B MoE variants, while Poolside released Laguna-M.1. This coordinated shift by multiple AI companies signals sustained competitive positioning through open-source distribution rather than closed APIs, contrasting with the API-first strategies of frontier labs and reflecting different go-to-market economics for smaller players.
Wall Street Journal via Techmeme
Researchers report that Zhipu AI's GLM-5.2 performs comparably to leading American models like Anthropic's Mythos and Fable 5 in cybersecurity applications, specifically on security bug-finding tasks. This signals competitive parity on a high-value capability and has prompted critics to question the effectiveness of current US export restrictions on Chinese open models.
Financial Times via Techmeme
Google told Meta around March that it couldn't fulfill all requested Gemini compute capacity, disrupting Meta's internal AI projects and reportedly prompting Meta's subsequent deals with alternative providers like Nebius. This concrete signal confirms compute remains the scarcest bottleneck in AI infrastructure and validates the economics driving hyperscaler capacity buildout and third-party infrastructure competition.
Tencent via Import AI
Tencent disclosed details on ARGUS, custom software used to trace, debug, and optimize large-scale training workloads, deployed across a 10,000+ GPU production cluster. The system has been used to manage 12,960-GPU MoE runs and 4,096-GPU video model training. The existence of such specialized infrastructure signals maturity in Tencent's training environment and broader sophistication in Chinese AI infrastructure buildout.
Epoch AI via Azeem Azhar, Exponential View
Data from Epoch AI shows a structural difference in how Chinese and American AI companies are building teams: Chinese AI labs hire talent with an average of 1.6 years of experience, compared to 5.5 years in US labs. This disparity has implications for training pipelines, retention, and competitive positioning in model development, and may reflect different cost structures and scaling strategies.
AI quarterly revenues now exceed quarterly CapEx depreciation, but cumulative revenues haven't covered historic investment
Azeem Azhar, Exponential View
According to Exponential View's reconstruction of the AI economy from the bottom up, quarterly AI revenues have crossed the threshold of quarterly CapEx depreciation for the first time. However, cumulative revenues have not yet covered the total capital invested historically, indicating the AI economy remains unprofitable at scale and raising questions about when the sector reaches sustainable unit economics.
