Short version: The instinct when you decide to self-serve an open model — DeepSeek V4 Flash, Qwen, Kimi K3 — is to open six GPU-cloud pricing pages and sort by hourly rate. Don't. Half of these products don't sell you the thing you think you're buying. Compare the product shape first, then let one number — utilization — decide between renting by the hour and paying by the token. (Prices below are reported mid-2026 figures; GPU pricing moves weekly, so verify on the provider's page before you commit.)

The six aren't the same shape#

CoreWeave is the one everyone name-drops and the one most founders should skip for this job. It sells node-scale infrastructure — there's no single-GPU SKU; you rent a full 8-GPU HGX node (reported ~$49/hr). It's built for training runs and cluster-scale inference, it's now public (CRWV), and it's the wrong tool for serving one 30B model. Renting eight H100s to run a model that fits on one is how you light money on fire.

Lambda is the straightforward answer: a single on-demand GPU VM (reported ~$3.99/hr for an H100 SXM). You get a box, you run vLLM yourself, that's it. No serverless, no scale-to-zero — a fixed hourly rate for a raw GPU. Simplest mental model in the group.

Nebius (public, NBIS; NVIDIA-backed) is a vertically integrated GPU cloud with on-demand (~$3.85/hr) plus a cheaper preemptible tier (~$2.15/hr). Preemptible is not serverless — your instance can be reclaimed — but for fault-tolerant batch or dev work it's the cheapest way to keep a real GPU under you.

RunPod is the most founder-friendly, on both axes. It has the cheapest raw GPU (reported ~$2.89/hr Secure, lower on Community Cloud) and true serverless: scale-to-zero workers, per-second billing, sub-2-second cold starts, no egress fees. If your traffic is spiky, this is where you don't pay for the 3am dead hours.

Together mostly sells you the serving, not the GPU. Its default path is a per-token API for popular open models — you may never touch a container. It also offers dedicated endpoints and raw Instant GPU Clusters (~$3.49/hr) if you want them, but the reason to pick Together is to skip the GPU entirely.

Baseten is the same idea, productized for production: managed model endpoints (package with the open-source Truss framework), billed per-GPU-minute with idle replicas scaled to zero for free. Not a raw-VM renter — a serving platform for teams without an infra person. We compared its neighbors in managed inference: Together vs Fireworks vs Baseten.

The only math that decides it#

Forget the hourly deltas for a second, because they're rounding error next to the real question: are you going to keep the GPU busy?

A pinned on-demand H100 runs roughly $1,800–2,900/month across the cheaper clouds. At DeepSeek V4 Flash's hosted price of $0.28 per million output tokens, that same budget buys well over 10 billion tokens — with zero ops, zero cold starts, and zero capacity risk. So the break-even isn't subtle:

Below roughly 6–12 billion tokens a month, a per-token API beats renting a dedicated GPU — not by a little. Above it, and only if you keep the GPU near fully utilized, owning the box wins.

And when your traffic is spiky — below ~40–50% utilization — scale-to-zero serverless beats a flat hourly rate, because you stop paying for the hours nobody is served. That's the RunPod-Serverless / Baseten case. The flat rental only pulls ahead once traffic is steady enough that the GPU stays hot and you're no longer paying a premium for elasticity you don't use. This is the same utilization logic behind Modal vs Replicate vs RunPod vs Baseten, one layer down the stack.

So what do you rent?#

The deeper comparison of the three integrated clouds — CoreWeave, Lambda, Nebius — lives in our GPU-cloud head-to-head. But the meta-point survives every price change on every one of these pages: rent by the hour only when the GPU stays hot. Everything else is a per-token API you haven't switched to yet.