Sunday, Nov 22, 2026

AI;DR
2026-11-22 · 13 stories- nvidiaNeoclouds are reselling Nvidia B300 capacity originally booked by Chinese-affiliated buyers after US rules blocked Chinese firms from renting advanced chips abroad. China represented about 20% of global GPU rental demand, pressuring B300 prices lower.
- chipsCurrent and former officials told The New York Times that China’s MSS has run a covert campaign to help Chinese companies obtain advanced AI chips. The effort highlights how chip controls are being challenged through espionage.
- OpenRouter found AtlasCloud, Novita and StreamLake forwarded prompts and outputs to MiniMax in China, affecting MiniMax and ERNIE 4.5 VL requests from September 23 to October 19. Research using 2 million canary markers found privacy flags ignored and prompt theft.+2
- UN states set 2028 as the target for regulating lethal autonomous weapons, but agreed to allow indirect human control rather than approval of every strike. Russia and the US resisted binding limits, leaving the instrument’s legal status undecided.+1
- regulationPandex puts legislation and top-court decisions from the EU, seven European countries and the European Court of Human Rights into one API and MCP server. It is built for search and AI agents.
- fundingThe IEA projects $4 trillion in global data center investment by 2030, while fewer than 100 major AI projects could need 50GW of power. Grid constraints and shortages of copper, cables and transformers may delay construction.
- amazonAmazon is nearing a $7.5 billion deal to acquire AI startup Decart, with talks advanced enough that an agreement could come as soon as next week. The purchase would help AWS compete with Nvidia for computing power.
- nvidiaDell, HPE, Lenovo and Supermicro are demanding substantially higher prices for AI servers months after contracts were signed. The increases are hitting trading firms that bought the equipment.
- The bank handling Anthropic’s IPO closed a position intended to stabilize the company’s stock. The move concerns the bank’s role in supporting trading around the offering.
- safetyColumbia University and CertiK introduced TensorLean to formally verify production GPU kernels at the PTX level using Lean. It checked 149 kernel instances, proved 83 correct and found 38 bugs in FlashAttention, DeepGEMM and other kernels.
- Nvidia researchers argue that huge context windows are slow, costly and error-prone when models do not need all their contents. In tests, periodic summaries and targeted keyword search usually matched full-context results at a fraction of the cost.
- China’s biggest tech companies are expected to double capital spending to $139 billion in 2026, while AI lab revenue accounts for less than 8% of that amount. Rhodium says this funding gap may matter more than chip restrictions.
- metaAWS Quick now has a Meta ads AI connector that lets advertisers create, manage and analyze campaigns in natural language. Users can combine Meta ad data with other sources to build dashboards, documents, slide decks and automated workflows.