Hey there. I saw you mentioning RunPod and honestly, thats what saved my skin when I was in a similar spot a few months ago. I spent ages trying to catch an instance on Lambda and it just never happened, it was like playing whack-a-mole with their availability. I finally gave up and rented a NVIDIA A100 80GB SXM4 through RunPod and it was a total breeze to get going. Since you mentioned being more of a dev than a devops person, you will probably love their pod templates. You basically just pick a vLLM or PyTorch template and it handles the driver mess for you. I remember being so stressed about the container not seeing the GPU but it worked right away... super satisfied with that experience. For something like DeepSeek-V3, you're gonna need a ton of VRAM, so if the A100s are eating your budget too fast, check out a multi-GPU setup with like 4x NVIDIA RTX 4090 24GB on their community cloud. Its cheaper and usually has way better availability than the high-end data center stuff. I've been running my prototypes there for a bit now and have zero complaints. Just make sure to use a persistent volume so you dont lose your model weights every time you spin down the pod to save some cash. It is basically the only way I have stayed under my own budget while still getting decent performance.

V3 uses a Multi-head Latent Attention architecture so VRAM efficiency is better, but at 671B params, you still need massive memory. For $500, a private cluster is out of reach. Check Vast.ai NVIDIA RTX 6000 Ada 48GB for the cheapest raw compute. Otherwise, just use DeepSeek DeepSeek-V3 API to hit that Friday deadline without a devops nightmare.