Overview
The week ending February 28 showed frontier AI firms competing through capital, government access, and public trust as much as model quality. OpenAI aligned itself with three infrastructure suppliers and the Pentagon; Anthropic's refusal to relax military-use limits produced consumer attention; open-model work exposed the engineering costs hidden by smaller files.
Developments
1. OpenAI joins capital, compute, and classified access
OpenAI announced $110 billion in investment from Amazon, Nvidia, and SoftBank at a $730 billion pre-money valuation, then disclosed an agreement to place its models on classified U.S. defense networks. The same week therefore expanded both the supply available to OpenAI and the institutions able to buy its services.
The combination strengthens OpenAI's position as an infrastructure intermediary, not only a model vendor. Cloud, chips, financing, and government procurement now reinforce one another; that alignment accelerates deployment while giving a small group of suppliers and public customers disproportionate influence over capacity and product priorities.
Sources: OpenAI financing · OpenAI defense agreement · New York Times
2. Military-use policy becomes a consumer acquisition channel
Claude reached No. 2 in the U.S. free iPhone chart after Anthropic's dispute with the Pentagon over mass surveillance and fully autonomous weapons. The ranking connected an enterprise policy position to a visible change in consumer distribution, even as government access shifted toward OpenAI.
Anthropic briefly converted exclusion from defense work into brand differentiation. That trade can support consumer growth but leaves the company exposed where federal procurement rules propagate through contractors, while OpenAI gains classified access at the cost of defending how its contractual safeguards operate in practice.
Sources: CNBC · TechCrunch
3. Washington turns data localization into a trade dispute
A February 18 State Department cable directed U.S. diplomats to identify and oppose foreign data-localization measures viewed as barriers to American cloud and AI services. Reporting on the cable this week placed data residency alongside tariffs and market access in U.S. technology diplomacy.
The policy raises the stakes for countries using localization to enforce privacy, security, or industrial policy. American providers gain diplomatic support for centralized cloud architectures, while foreign regulators face pressure to defend rules whose practical effect is to keep training data, prompts, or model outputs inside national borders.
Sources: Reuters · TechCrunch
4. Qwen quantization exposes a speed penalty below the file-size gain
Unsloth's Qwen 3.5 research covered more than 150 KL-divergence tests and roughly 9 terabytes of published GGUF artifacts. Some lower-bit methods reduced model size while slowing inference by 5% to 10%, and several tensor groups proved unusually sensitive to quantization.
Those measurements make precision a deployment variable with competing effects, not a one-directional compression dial. The winning configuration changes with memory limits, GPU architecture, throughput targets, and task sensitivity, weakening broad claims that a smaller local model is automatically cheaper to serve.
Sources: Unsloth benchmarks · Published artifacts