Overview
OpenAI closed February by securing both a classified-network agreement and $110 billion in new capital from infrastructure partners. Anthropic converted the same defense dispute into consumer attention, while Qwen quantization research showed that smaller model files can carry non-obvious speed and quality costs.
Developments
1. OpenAI defines boundaries for a classified Pentagon deployment
OpenAI said its agreement would make AI models available on classified U.S. defense networks through a cloud deployment. The published structure leaves its safety stack under company control, places cleared OpenAI personnel in the operating loop, and bars mass domestic surveillance and direction of autonomous weapon systems.
The arrangement gives a private model provider continuing authority inside a classified environment, an unusual division of responsibility for defense software. Its durability turns on whether OpenAI can observe prohibited uses and exercise its termination rights without interrupting missions that have come to rely on the service.
Sources: OpenAI agreement · New York Times
2. Claude reaches second place in the U.S. free iPhone chart
Anthropic's Claude rose to No. 2 in Apple's U.S. free-app ranking after the Pentagon dispute, CNBC and TechCrunch reported. A procurement conflict had crossed into consumer distribution, giving Anthropic a burst of attention tied directly to its limits on military AI use.
The ranking captures download velocity, not audience size or retention. Anthropic gained a low-cost acquisition event from a government confrontation, but the commercial benefit fades quickly unless those downloads become recurring usage and paid subscriptions after the policy controversy leaves the front page.
Sources: CNBC · TechCrunch
3. OpenAI raises $110 billion from three infrastructure partners
OpenAI announced $110 billion in investment at a $730 billion pre-money valuation, with Amazon, Nvidia, and SoftBank supplying the capital. The investors also occupy critical positions in cloud capacity and accelerator supply, binding the financing to the infrastructure OpenAI intends to buy.
The round gives OpenAI exceptional purchasing power while concentrating more of its expansion among three counterparties. Amazon and Nvidia can benefit twice, first as investors and again as suppliers, which may narrow OpenAI's room to shift workloads or pricing when infrastructure interests diverge.
Sources: OpenAI · TechCrunch
4. Unsloth publishes a large Qwen 3.5 quantization study
Unsloth published more than 150 KL-divergence tests and roughly 9 terabytes of Qwen 3.5 GGUF research artifacts. The experiments identify tensor groups that degrade sharply under lower precision and report a 5% to 10% slowdown for some quantization methods despite their smaller files.
The slowdown breaks the simple assumption that fewer bits always yield faster local inference. Quantizer choice now becomes a hardware-specific trade among memory capacity, token throughput, and downstream quality; a method that fits an extra model into memory can also lengthen every response or damage a sensitive capability.
Sources: Unsloth benchmarks · Unsloth research artifacts