Newsletter Archive

As inference workloads dominate real-world AI deployments, specialized chip startups are carving out niches that train-optimized giants like Nvidia can't easily defend — lowering costs and latency for production models.

Theregister · 6 min read Hardware

AI companies' relentless hunt for training data is flooding the market with low-quality junk that threatens physical AI and world models. The shift from internet-scale LLM training to embodied AI demands rich, multifaceted real-world data that can't be mass-produced, and the glut of synthetic/inadequate datasets is becoming a crisis.

Fortune · 4 min read Industry

A data center expert warns that massive power demands from AI infrastructure buildouts could exceed grid capacity and trigger rolling blackouts in some regions, raising serious questions about whether electricity infrastructure can support the gigawatt-scale facilities being rushed into production.

24/7 Wall St · 3 min read Industry

Nvidia's dominance in China's AI chip market is eroding as US export restrictions on advanced GPUs drive Chinese companies toward homegrown alternatives like Huawei's Ascend and domestic startups, reshaping the global AI hardware landscape.

Cryptobriefing · 3 min read Hardware

Applied Materials is acquiring NEXX Systems from ASMPT Limited to bolster advanced packaging capabilities for AI processors, addressing growing demand for the specialized infrastructure needed to connect and cool high-performance AI chips.

Quiverquant · 9 min read Hardware

TokenArena benchmarks 78 AI inference endpoints across 12 model families using five core axes (speed, latency, price, context, quality) plus modeled energy, revealing the same model differs wildly across providers—up to 12.5 points in accuracy, 6.2x in joules per correct answer, and dramatically reordering leaderboards depending on workload (chat vs. retrieval vs. reasoning).

arXiv AI · 4 min (abstract) Research

AgentFloor benchmark across 16 open-weight models (0.27B–32B) and GPT-5 reveals a clear capability boundary: small and mid-sized models handle the majority of production agent workflows (short, structured tool use), while frontier models maintain advantage only on long-horizon planning with persistent constraints. Open-weight models can match GPT-5 on the benchmark at lower cost/latency.

arXiv AI · 3 min (abstract) Research

A decade after Geoffrey Hinton predicted AI would obliterate radiology, the profession has grown 10% and radiologist salaries now reach $571K as demand surges. Even Hinton later walked back his claim, acknowledging radiologists will augment rather than be replaced by AI.

Fortune · 5 min read Community