Best cloud provider for hosting DeepSeek API?

Question

Which cloud provider is going to give me the best performance and cost balance for self-hosting the DeepSeek-V3 API for a production app?

Im currently torn between AWS and RunPod, leaning towards RunPod because of the raw price difference for GPU rental, but I am really worried about uptime since this is going to be a customer-facing tool.

Here is the setup. I am building a customer support routing tool for a client, and we need to launch in about six weeks. Because of their data privacy policies we cannot use the public DeepSeek API, so we have to host the weights ourselves. Our budget is pretty tight, ideally under 800 dollars a month for the hosting side.

Right now I am looking at these options:

AWS: We could use SageMaker or just spin up some EC2 instances. The big pro is that the rest of our backend is already there, and we know the uptime is solid. The con is that GPU availability is a nightmare right now and the costs might push us way over budget.

RunPod: The pricing is incredibly cheap compared to AWS and they actually have GPUs available. But I keep reading stories about nodes randomly dying or storage volumes disconnecting, which scares me for a production release.

Lambda Labs: A middle ground, but every time I check their cloud console, the instances I need are completely sold out.

Is the cost saving on RunPod actually worth the potential stability headaches, or should we just suck it up and pay the AWS premium for peace of mind?

SteveWem · Accepted Answer

Late to the party here but had to jump in because we just went through this exact headache last month! Running a quantized DeepSeek on AWS almost gave our CFO a heart attack. Renting a couple of NVIDIA RTX 6000 Ada 48GB GPUs on a dedicated provider ended up being the sweet spot and it has been absolutely fantastic! My quick tip for that budget: you absolutely must use vLLM with a 4-bit AWQ quant to fit your memory requirements without going broke. Also, look into dedicated bare-metal GPU hosts like Hostkey instead of marketplaces. They offer real SLAs, which completely solves the RunPod reliability anxiety!

voxrrphtud · Answer

Regarding what #1 said about "Is the cost saving on RunPod actually worth the stability headaches" - honestly, absolutely not for a client-facing launch. I had issues with their network storage mounts just disappearing mid-query, which is a total dealbreaker. Here is my take on the options right now:

AWS: The uptime is unmatched, but unfortunately, spinning up even a single NVIDIA A100 80GB SXM4 is going to blow right through your 800 dollar budget. The pricing is just not as good as expected for small teams.

Dedicated hosters: You get better stability than RunPod and lower prices than AWS. But unfortunately, GPU availability is still pretty terrible across the board right now. If you can, try to squeeze the budget and go with AWS, or look at reserved instances on smaller specialized clouds. Dont risk RunPod for a client.

care about people around you_z · Answer

> Is the cost saving on RunPod actually worth the potential stability headaches, or should we just suck it up and pay the AWS premium for peace of mind? Honestly, RunPod is amazing for dev, but for a production client launch in six weeks, those random node drops will give you nightmares! Ive been using them for a year and love the savings, but you definitely get what you pay for. Since your budget is tight at 800 bucks, AWS is going to kill you on costs. You wont even get a single decent GPU for that. My advice? Go with RunPod but spin up two smaller instances behind a basic load balancer to handle the stability issues. You can rent a couple of NVIDIA RTX 6000 Ada 48GB GPUs for around 75 cents an hour each. That easily fits your budget and gives you redundancy if one node randomly dies. Just make sure to script the auto-recovery! It works like a charm and saves you thousands compared to AWS.