AI Roundup: April 5, 2026

1. DeepSeek V4 Will Run Entirely on Huawei AI Chips

DeepSeek. China’s DeepSeek confirmed that its upcoming V4 model, a one-trillion-parameter mixture-of-experts architecture with roughly 37 billion active parameters per inference pass, will run entirely on Huawei-designed AI chips. DeepSeek spent months collaborating with Huawei and Cambricon Technologies to rewrite core model components for non-NVIDIA hardware. Chinese tech giants including Alibaba, ByteDance, and Tencent have placed bulk preorders for Huawei’s chips totaling hundreds of thousands of units, and DeepSeek has reportedly denied early access to both NVIDIA and AMD. A mid-to-late April launch is expected. Source

2. Apple Approves Tiny Corp TinyGPU Driver for NVIDIA and AMD eGPUs on Apple Silicon

Apple / Tiny Corp. Apple has signed a third-party GPU driver built by Tiny Corp that enables NVIDIA and AMD external GPUs to run on Apple Silicon Macs over Thunderbolt/USB4. The TinyGPU driver extension is specifically designed for AI research workloads, not display rendering or gaming. This marks the first time macOS has supported third-party discrete GPUs on Apple Silicon through an officially sanctioned mechanism, removing the need to disable System Integrity Protection. AI researchers can now pair high-end discrete GPUs with Apple hardware for model training and inference. Source

3. Iran Strikes Leave AWS Availability Zones Offline in Bahrain and Dubai

AWS. Continuing strikes by Iran’s Islamic Revolutionary Guard Corps have knocked key AWS availability zones in Bahrain and Dubai offline, forcing Amazon to urge customers to shift workloads to other regions. A drone struck a Bahrain telecom facility hosting AWS infrastructure on April 1-2, following earlier attacks on two UAE data centers in March. Outages affected Abu Dhabi Commercial Bank, Emirates NBD, First Abu Dhabi Bank, ride-hailing platform Careem, and Snowflake. The incidents represent the first publicly confirmed military attacks on a hyperscale cloud provider, fundamentally altering how geopolitical risk must be factored into multi-region cloud architecture. Source

4. Wharton Study Finds “Cognitive Surrender” — Users Follow Wrong AI Advice 80% of the Time

Wharton. A study by researchers Steven Shaw and Gideon Nave at the University of Pennsylvania tested 1,372 participants across 9,593 trials and found that users followed AI advice 92.7% of the time when correct but still complied 79.8% of the time when the AI gave demonstrably wrong answers. The researchers coined the term “cognitive surrender” and proposed a Tri-System Theory extending dual-process cognitive models by adding System 3 — artificial cognition that overrides both intuition and deliberation. Participants with higher AI trust and lower need for cognition showed the greatest surrender effect, raising concerns for high-stakes domains like healthcare and education where uncritical AI reliance could cause harm. Source