Newsletter Archive

Nvidia's production costs now depend 90% on Asian suppliers, up from 65% a year ago, intensifying supply-chain vulnerability as demand for AI chips and physical AI infrastructure accelerates.

Tom's Hardware · 1 min read Hardware

Anthropic is launching a new enterprise AI services company in partnership with investment firms Blackstone and Hellman & Friedman plus Goldman Sachs, marking a major expansion beyond model development into managed services and consulting for large organizations.

Anthropic · 5 min read Industry

Elon Musk is suing Sam Altman and OpenAI, claiming they abandoned the nonprofit mission to prioritize profits—seeking removal of leadership and up to $150B in damages. The trial began in late April with Musk testifying across three days; week two features expert witnesses including Stuart Russell and Greg Brockman.

The Verge AI · 3 min read Policy

Amazon SageMaker AI now lets you define a prioritized list of instance types for inference endpoints, automatically falling back to available capacity during creation, scaling, and downscaling without manual retries. Solves the persistent problem of GPU capacity constraints blocking LLM deployments.

Amazon AWS ML · 13 min read Tools

Anthropic launched a $1.5 billion venture with Goldman Sachs, Blackstone, and others to embed Claude into private equity-backed companies. The fund targets enterprise adoption at scale, positioning Anthropic to compete directly with OpenAI across Fortune 500 and mid-market firms.

CNBC Tech · 1 min read Industry

Cerebras, an AI chip maker competing with NVIDIA, plans to raise $3.5B in an IPO at a $26.6B valuation, marking a significant move in the race for AI compute dominance as alternatives to NVIDIA GPUs gain traction.

SiliconANGLE · 5 min read Market

Meta is securing a roughly $13 billion financing package (via Morgan Stanley and JPMorgan) for an AI data center in El Paso, Texas—a signal that Big Tech is turning to debt markets to fund the massive infrastructure buildout driving modern LLM development.

Bloomberg Tech · 1 min read Industry

Researchers developed a simple perplexity-differencing method to expose hidden behaviors injected via finetuning in LLMs, testing it on 76 model organisms from 0.5B to 70B parameters including backdoored and adversarially trained models. The technique works by ranking completions by perplexity gaps between reference and finetuned models, surfacing finetuning objectives without requiring access to model internals.

arXiv NLP · 3 min (abstract) Research