Curious about one thing: are you gonna use quantized weights or the full version? Honestly, I've been really happy with Fireworks AI for serverless. It works well and I've had no complaints! Basically, you pay per token instead of renting a whole NVIDIA H100 80GB Tensor Core GPU rig, so it handles the VRAM scaling for you. Ngl, let me know your latency needs!

So basically the consensus is RunPod or Fireworks AI. Renting dedicated NVIDIA A100 80GB nodes is unfortunately usually a money pit... maybe try DeepInfra? Value is better there. gl

Saved for later, ty!

Yo! I've been tinkering with R1 and ngl, it's a total beast!! For your situation, I recommend: 1. Go with RunPod, they're amazing for casual testing cuz you're not paying a fortune for idle stuff.
2. Check out Lambda Labs for top-tier reliability and professional standards. Honestly, just be super careful with memory allocation... your VRAM usage will spike fast! Its fantastic but stay cautious with you're setup. gl!

I spent a few weeks trying to find a stable middle ground for these massive models because I really value uptime over saving a few cents. Started with some cheaper spot instances elsewhere but honestly, getting kicked off mid-inference was a nightmare for my sanity. I ended up testing Vultr NVIDIA H100 PCIe and found their stability much better than the smaller specialized shops. Its not the absolute cheapest, but I dont have to worry about the instance disappearing. I also messed around with Groq LPU Inference Engine for a bit... the speed is actually insane if you can get into their early access, but it feels a bit different than a traditional stack. Quick tip: double check your egress costs. Some providers lure you in with cheap hourly rates but then hit you hard when you try to move data or weights back and forth.