RTX 3090 tempts local AI builders

The sharp take: local inference is becoming a serious cost hedge, but the frontier coding models still win when reasoning quality, speed, and long-context reliability matter.

• –36GB VRAM can run useful quantized 30B-35B-class coding models, especially MoE models like Qwen3.5 35B-A3B, but context size and quantization choices still define the real experience.
• –Dual-GPU setups look attractive on paper, yet mixed 3090/4070 rigs add power, cooling, PCIe, offload, and model-sharding complexity that cloud subscriptions hide.
• –The strongest workflow is hybrid: local models for cheap routing, classification, boilerplate, and subagent work; paid Sonnet/Opus-class models for hard planning and final reasoning.
• –Buying more VRAM is not future-proof in a simple way; models are getting more efficient, but developer expectations for context, tool use, and parallel agents are rising just as fast.