Newsletter Archive

AMD beat Q2 revenue guidance by $700M to $11.2B, driven by data center AI chip demand, challenging Nvidia's dominance as the primary alternative for training and inference infrastructure.

Bloomberg Tech · 1 min read Market

OpenAI launched new realtime voice models in its API with reasoning, translation, and transcription capabilities, enabling developers to build more natural voice-based AI applications without the latency of previous approaches.

OpenAI · 1 min read Tools

SpaceX plans to invest at least $55 billion into Terafab, an AI chip fab in Texas, with potential expansion to $119 billion total. The facility aims to produce chips supporting 200+ gigawatts of annual compute capacity, marking Musk's direct entry into the AI silicon market competing with TSMC and domestic foundries.

The Verge AI · 2 min read Hardware

Mira Murati's deposition in Musk v. Altman exposes what happened behind closed doors during Sam Altman's November 2023 ouster from OpenAI, including board communications and the chaotic 72-hour power struggle that the public only saw fragments of in real time.

The Verge AI · 9 min read Industry

Anthropic introduces natural language autoencoders that convert Claude's internal activations into readable text, enabling direct observation of model reasoning. This interpretability advance lets researchers identify and potentially steer specific model behaviors without retraining.

Anthropic · 16 min read Research

Elon Musk is leasing Anthropic his world's-largest supercomputer for $3-4B annually, generating 2.5B+ in profit margins. The deal—announced just three months after Musk called Anthropic 'evil'—doubles as a revenue anchor for SpaceX's upcoming IPO and marks his shift from owning xAI to becoming the infrastructure landlord powering a major AI rival.

Fortune · 5 min read Industry

OpenAI is rolling out GPT-5.5-Cyber, a version with fewer guardrails, to vetted cyber defenders through its Trusted Access program. The model matches Anthropic's Mythos at finding and exploiting software bugs, sparking White House concerns about keeping such capabilities from bad actors.

Axios · 3 min read Industry

LightSeek Foundation releases TokenSpeed, an MIT-licensed open-source inference engine optimized for agentic AI workloads like coding agents. Unlike traditional LLM serving, agentic systems need to handle 50K+ token contexts and multi-turn conversations while balancing per-GPU throughput (TPM) and per-user responsiveness (TPS) — a problem TokenSpeed's five-subsystem architecture directly targets.

MarkTechPost · 5 min read Tools