Daily AI RoundupsMonday, 13 July 2026

The 13 Jul 2026 Calathea roundup leads with Helsing raising $1.8B at an $18B valuation in Europe’s largest defense-AI round, then tracks Meta’s planned $50B-plus Hyperion data center supercluster, NVIDIA’s revenue-sharing and credit-support model for AI cloud buildouts, and TSMC’s record Q2 revenue on AI chip demand. It also covers Ben Thompson’s Stratechery analysis of Apple’s OpenAI lawsuit, George Hotz arguing that useful LLMs do not mean frontier labs will capture AI’s value, Mach-Mind-4-Flash’s post-training efficiency result, Soofi S 30B-A3B’s sovereign open-source MoE hybrid model, and MIT’s auditing method for harmful model capabilities.

AI infrastructure
defense AI
AI chips
frontier labs
open-source AI
MoE models
AI safety

Stories in this roundup

Helsing raises $1.8B at $18B valuation in Europe’s largest defense-AI round, with investor demand significantly exceeding allocationCNBC
Meta’s Louisiana Hyperion data center supercluster to cost over $50B and reach 5GWCNBC
NVIDIA unveils revenue-sharing and credit-support model for AI cloud buildouts, with partners deploying up to 170,000 GPUsNVIDIA
TSMC posts record Q2 revenue, jumping 36% YoY on AI chip demandReuters
On Stratechery, Ben Thompson argues Apple’s trade-secret lawsuit against OpenAI feels like lashing out and masks deeper strategic problemsStratechery
George Hotz argues LLMs are useful, but frontier-lab hype overstates moats, value capture, and light-cone narrativesgeohot blog
Mach-Mind-4-Flash: a 35B MoE agentic model with 3B activated parameters achieves strong agent performance via post-training optimization alone, without scaling pre-training computearXiv
Soofi S 30B-A3B: a sovereign open-source MoE hybrid Mamba-Transformer foundation model for German and English activates only 3B of 30B parametersarXiv
MIT researchers develop auditing method to test generative models for harmful capabilities without prompting illegal outputsMIT News

Back to all roundups