Honestly, I've spent years jumping between providers and for a beast like DeepSeek-V3, RunPod GPU Cloud is basically unbeatable right now. SageMaker is cool but sooo expensive if you're on a budget. I've been running MoE models on Lambda Labs GPU Cloud too, and while the uptime is usually fine, RunPod feels more flexible for real-time chat. Ngl, the price-to-performance on their NVIDIA H100 Tensor Core GPU instances is amazing for low latency! gl!

Seconding the recommendation above. RunPod is great, but honestly, if youre worried about uptime for a production app, you might wanna look at Vast.ai or CoreWeave GPU Cloud for more enterprise-grade reliability. V3 is a beast to host because of that MoE structure, right? I've been really happy with DigitalOcean GPU Droplets lately too. Theyre stable, and the pricing is lowkey better than AWS for scaling. Just be careful with spot instances if you need low latency, you know? gl!

Bookmarked, thanks!

> I’m looking for a cloud provider that offers the best price-to-performance ratio for inference. Tbh if ur doing market research on this, look at Oracle Cloud Infrastructure. Everyone sleeps on OCI but their RDMA networking is CRITICAL for MoE models like V3 to hit those low latency targets without costing a fortune. If you want stability without the SageMaker tax, check out Together AI. They’ve optimized the inference kernels specifically for V3's architecture, so the price-to-performance is lowkey better than renting raw GPUs and DIYing it (at least that's what worked for me). Don't overpay for the brand name when specialized providers are faster.

Yeah, I definitely agree with the point about optimized kernels being a game changer, but as someone who has been managing high-parameter deployments for years, there is a serious case for going the DIY route with dedicated bare metal if you want to maximize your margins. If you have the dev ops skills to handle the setup, you can often beat the pricing of managed providers once your volume scales. Here are a few options I have used for self-managed V3 deployments that offer a different vibe from the usual suspects:

• FluidStack: I have found them to be the sweet spot for sourcing high-end H100s or A100s across different data centers. Their reliability is a step up from the purely consumer-grade clouds, and you get dedicated access which is vital for MoE stability.
• LeaseWeb: This is for the true veterans who want zero hypervisor overhead. Running V3 on bare metal means you get every ounce of performance from the VRAM and interconnects, which helps a lot with that real-time chat latency you are after.
• Vultr GPU: Their orchestration is actually pretty mature now. It is a bit more enterprise-ready than the hobbyist clouds but significantly cheaper than the SageMaker tax. Basically, if you can tune vLLM or SGLang yourself, the DIY approach is way more rewarding long-term. Just depends if you have the time to babysit the instances. Anyway, good luck with the launch!

^ This. Also, I just saw this thread and felt like I had to jump in because I've been down the managed cloud rabbit hole and it's honestly a total trap. Last year I was running a similar MoE setup on one of those shiny providers everyone loves, and a single unexpected maintenance window basically bricked my production app for six hours. It was brutal, ngl. I would suggest taking a step back and looking at the DIY route. It might be more work, but you'll thank yourself when you arent paying a 40% markup for a dashboard you barely use. Be careful though, the networking is where they get you if you dont plan it right.

• Just get some dedicated machines from Hetzner Dedicated Server.
• Go with the bare metal options from OVHcloud Bare Metal.
• Look into any of the local boutique data centers near you. Honestly, once you get your own environment tuned, the stability is so much better for V3. You just have to be willing to get your hands dirty with the orchestration and ignore the hype around these auto-scaling platforms. It's more stable in the long run imo...