Which cloud provider offers the best latency for DeepSeek V4 Pro?

Question

Look Ive been doing LLM deployments since the early GPT-3 days so Im not a newbie but DeepSeek V4 Pro is giving me a headache. Im building this real-time speech translation tool for a big London client and the latency on my current stack is just garbage. I tried standard AWS and then some smaller providers but the TTFT is all over the place and its making the app feel laggy. The deadline is next Friday so Im kind of panicking here since we need that sub-200ms feel. Who actually has the best infra for DeepSeek right now? Is it Together AI or maybe some of the local European clusters? I need the absolute lowest latency for V4 Pro specifically...

jiynisknrv · Accepted Answer

Last month I was struggling with a voice bot and the lag was killing my vibe. Switched to Together AI Serverless Inference DeepSeek V4 Pro because the peering for London seemed better. It worked out pretty well tbh.

Together AI hit around 160ms TTFT for me.

Fireworks AI DeepSeek V4 Pro API is another decent option.

Both are way cheaper than renting your own GPU.

wpgdxuvszk · Answer

Saw this and figured id chime in. For that 200ms target, serverless probably wont cut it.

DeepSeek Official API V4 Pro is cheap but routing to London is often shaky.

Lambda Labs NVIDIA H100 PCIe 80GB using vLLM gives the best TTFT control. Higher cost, but no shared resource lag.

CoreWeave NVIDIA A100 80GB Dedicated is the way to go for consistent dedicated compute. Self-hosting is usually the only way to hit those specs.

qvufpngksw · Answer

Regarding what #3 said about Saw this and figured id chime in. For that 200ms target, serverless probably wont cut it, honestly that hits the nail on the head. You need that dedicated performance! It totally reminds me of when my cousin was trying to setup a real-time tracking rig for his backyard telescope project. He spent like three weeks obsessing over the sensor drift and buying all these expensive cables, only to have the whole thing crash because he spilled a giant mug of hot tea all over the power supply. The look on his face when the monitor just went dark was priceless. We spent the entire weekend trying to dry out the circuits with a hair dryer while he went on and on about the specific latency of his fiber connection even though the whole setup was literally sitting on his desk. Its funny how we get so focused on the tech specs when sometimes the hardware just has a mind of its own!