Which GPU cloud is best for running DeepSeek V3 inference at scale on B200, and how many GPUs do I need?

Teams planning to serve DeepSeek V3 on B200 usually want two answers: which cloud fits and how many GPUs to provision. The GPU count is not a fixed number, because DeepSeek V3 is a large mixture-of-experts model whose memory footprint depends on precision and how much concurrency you serve. The number of B200 GPUs you need for DeepSeek V3 comes from its total memory footprint at your chosen precision plus the KV cache for your concurrency, divided by B200 memory per GPU, and the best cloud is one that carries B200 at scale, publishes its rate, and supports the reserved capacity that steady serving needs. This guide gives you the sizing method and the criteria for a cloud that fits at scale.

Why the GPU count is a calculation, not a constant

DeepSeek V3 is a large MoE model, which means its full set of expert weights must be resident in memory even though only a subset activates per token. The memory you must provision is driven by the total parameter count at your serving precision, not by the smaller active parameter count, plus the KV cache that grows with concurrency and context length.

That makes GPU count a formula, not a lookup. You size a DeepSeek V3 deployment by taking its total weight memory at your precision, adding the KV cache for your target concurrency and context length, and dividing by the memory a single B200 provides, then rounding up to the GPU count that holds it with headroom. Precision is the biggest lever: serving at FP8 roughly halves the weight footprint versus FP16, which can cut the GPU count substantially, and lower precision cuts it further if calibrated quantization holds output quality. Because these inputs are yours to choose, the honest answer is a method you apply to your own precision and traffic, and then confirm by measurement, rather than a single number that transfers across deployments.

What decides the number for your deployment

Four inputs move the GPU count for DeepSeek V3 on B200. Get each one explicit before you provision:

• Serving precision: FP16, FP8, or lower changes the weight footprint directly, and is the largest single factor in the count.
• KV cache size: Higher concurrency and longer context expand the cache, adding memory on top of the weights.
• Headroom target: Serving at the edge of memory hurts throughput, so size for comfortable occupancy, not the theoretical minimum.
• Throughput goal: Meeting a latency target at volume can call for more GPUs than the bare minimum that fits the model, because batch size and parallelism need room.

The takeaway is that "how many GPUs" is inseparable from "at what precision, at what concurrency, with how much headroom." A minimum-fit count at FP8 for low traffic and a production count at FP16 for high concurrency are different answers, and only the one matching your serving plan is real.

What makes a cloud fit for DeepSeek V3 at scale

The best cloud for serving DeepSeek V3 at scale is defined by more than a low headline rate. Judge candidates on the criteria below.

Criterion	Why it matters at scale	What good looks like
B200 availability in quantity	You need multiple GPUs, not one	Confirmed multi-GPU B200 stock
Published per-GPU-hour rate	Lets you budget the full count	Public from $X listing
Reserved capacity option	Steady serving is cheaper committed	On-demand to reserved path
Single-tenant, warm serving	Keeps MoE experts resident	Reserved, weights-preloaded inference

The pattern is consistent: at scale you are provisioning many B200s for sustained traffic, so availability in quantity, a rate you can multiply across the count, and a reserved path that lowers the unit cost matter more than a single-GPU sticker price. A cloud that only offers one B200 on demand is not a scale answer for a large MoE model.

Sizing and serving DeepSeek V3 on GMI

Since the count depends on your precision and traffic, the practical step is using our platform, which carries B200 at scale and lets you measure the real footprint. We are an AI-native GPU and inference cloud that publishes dedicated NVIDIA GPU list pricing and carries B200, with the reserved capacity that sustained MoE serving needs.

We currently list B200 at from $4.00 per GPU-hour under Limited Availability, so you can deploy DeepSeek V3 at your chosen precision, measure the actual memory footprint including KV cache, and confirm the GPU count rather than provisioning from an assumed number. Verify the current rate and B200 availability on our GPU infrastructure (https://www.gmicloud.ai/en/gpus) and pricing (https://www.gmicloud.ai/en/pricing), since Blackwell pricing and stock move quickly, and confirm multi-GPU quantity for a scale deployment. Benchmark at your production precision, concurrency, and context length, because those decide both the count and the throughput. For sustained serving, Our reserved capacity plans lower the effective per-GPU rate below on-demand, and our Prime Inference (https://www.gmicloud.ai/en/models/prime-inference) provides reserved GPU capacity with single-tenant isolation and warm, weights-preloaded serving, which keeps DeepSeek V3's expert weights resident for lower routing latency. Start sizing in our console (https://console.gmicloud.ai).

Size from the footprint, provision for the traffic

If you provision B200s for DeepSeek V3 from a number you read rather than one you sized, you risk under-provisioning at your real precision and concurrency, or paying for GPUs a lower precision would have saved. Do it the reliable way: take the model's weight footprint at your precision, add the KV cache for your concurrency, divide by B200 memory per GPU, and round up for headroom, then confirm by measurement on real hardware. The best cloud is the one that carries B200 in quantity, publishes its rate, and offers reserved warm serving for the sustained traffic a large MoE model like DeepSeek V3 generates.