AI news digest — July 30, 2026
10 items, each with its source.
Microsoft Azure crosses $100B, Copilot hits 30M seats
Microsoft reported Q4 FY26 revenue of $90.0B, with Azure and other cloud services growing 43% year-over-year. CEO Satya Nadella announced Azure revenue exceeded $100B for the full fiscal year and Microsoft 365 Copilot crossed 30 million paid seats. This massive pre-committed enterprise backlog signals durable AI infrastructure demand.
aiweekly.coMeta raises 2026 AI capex floor to $130B
Meta reported Q2 2026 revenue of $60.8B and narrowed its full-year capital expenditure range upward to $130B-$145B. The increase is driven by massive AI datacenter buildouts and higher HBM memory chip prices. This massive infrastructure commitment solidifies Meta's position in the open-weight AI race.
aiweekly.coSamsung Q2 operating profit soars 19-fold on AI boom
Samsung Electronics posted a record quarterly operating profit, up 1,814% year-over-year, driven by robust AI-related demand for HBM and conventional DRAM. The company cited exponential growth in demand for general-purpose computing alongside AI infrastructure. This confirms that the AI infrastructure buildout is continuing to pay hardware suppliers handsomely.
cnbc.comPoolside Laguna S 2.1 beats DeepSeek V4 Pro Max
Poolside released Laguna S 2.1, a 118B parameter MoE model with 8B active parameters that outperforms DeepSeek V4 Pro Max on coding benchmarks. The model is positioned as a major Western competitor in open-weight coding models. This proves that frontier coding quality does not strictly require ever-larger parameter counts.
aitoolsrecap.comFrontier agents fail at independent AI research tasks
A new "Shadow Evaluations" paper tested frontier AI agents on central research questions from unpublished NeurIPS 2026 submissions. While the agents completed engineering work autonomously, they failed at actual research due to poor judgment, uncreative problem-solving, and instruction drift. This highlights the current limitations of AI agents in performing novel scientific research.
aiweekly.coRuflo MCP bridge flaw scores maximum CVSS 10.0
Noma Labs disclosed CVE-2026-59726, a critical vulnerability in the open-source AI agent platform Ruflo that exposed 233 tools. The flaw allowed unauthenticated remote code execution and exfiltration of LLM API keys via a single HTTP POST. This maximum-severity bug underscores the growing security risks in rapidly deployed AI agent frameworks.
aiweekly.coCoRT adds token-level credit to GRPO for LLMs
Researchers introduced CoRT, a method that rescores sampled responses using tokenwise log-likelihood contrasts to improve GRPO training. The technique achieved an average 4.4-point gain across instruction-tuned models without requiring an auxiliary scorer. This advancement provides a more efficient way to assign token-level credit during reinforcement learning for LLMs.
aiweekly.coSkillRise trains single agent policy across multiple environments
A joint research team developed SkillRise, which trains a single policy to alternate between solving tasks and curating an evolving skill document. The method achieved state-of-the-art Pass@1 rates across ALFWorld, WebShop, and ScienceWorld benchmarks. This unified approach demonstrates how agents can continuously evolve their skills across diverse environments.
aiweekly.coMeta's HumanCLAW tests VLMs on embodied physical tasks
Meta released HumanCLAW, a new benchmark for evaluating vision-language models in embodied physical environments. The evaluation found that top VLMs can currently solve only 16.8% of embodied SIT tasks. This significant failure rate highlights the ongoing gap between visual recognition capabilities and actual physical reasoning in robotics.
aiweekly.coTurboVLA drops LLM to run at 32 Hz on 4090
Researchers introduced TurboVLA, a new vision-language-action model architecture that completely removes the traditional LLM backbone. This optimization allows the model to run at a highly responsive 32 Hz on a single consumer-grade RTX 4090 GPU. The breakthrough enables high-speed robotic control without relying on massive compute clusters.
aiweekly.coFeed, daily deep-dive and bytes — readable offline, with push alerts for the topics you follow.