GPU-Powered VDI & Remote Desktops 2026 — vGPU vs Passthrough, Users per GPU & Server Sizing

The most common GPU-server question we hear is no longer about training — it is about desktops. Engineering teams want their CAD and 3D applications on thin clients in branch offices or factory floors; OT teams want Windows HMI sessions virtualized instead of bolted to aging PCs beside every machine; IT wants the data to never leave the building. All three converge on the same architecture: a GPU server running virtual desktops, with one NVIDIA license per concurrent user. This guide sizes that server — how many users per GPU, which sharing model, and what the software really costs in 2026.

Three ways to share a GPU — and when each is right

Every GPU-virtualization decision starts with the sharing model. NVIDIA's official licensing FAQ is blunt: a passthrough deployment is enforced through the EULA rather than per-CCU software, which is why 1:1 passthrough is the license-free default — but it is also the least dense option, and workstation-grade features still need an RTX Virtual Workstation entitlement.

MIG and vGPU cover different jobs: MIG partitions an A100/H100 for compute tenants, while vGPU multiplexes a graphics card for desktop users — our earlier MIG/vGPU deep-dive covers the node-level detail. The desktop choice is therefore vPC vs vWS, and it is driven almost entirely by the application list, not by the GPU.

Users per GPU: profile math that survives contact with reality

Density starts from the vGPU profile: vPC desktops commonly run 2–4 GB profiles and vWS workstations 8–16 GB, so a 48 GB card has a framebuffer ceiling of roughly 24 desktop profiles or 3–6 workstation profiles. NVIDIA's own sizing guidance adds two caveats: concurrent-vGPU limits per GPU, and the fact that time-slicing shares the same SMs — pack 24 users onto one L40S and every user feels the contention at login and in video. A defensible planning number is 8–12 vPC users or 3–4 vWS users per 48 GB Ada GPU for fluid interactive work.

Plan 1.5–2× the CPU and RAM you think you need: every vPC desktop wants 2 vCPU and 4 GB, every vWS session 4–8 vCPU and 16–32 GB. A 2-socket server with 1 TB RAM typically hosts 4× 48 GB GPUs — 32–48 vPC users or 12–16 vWS users per 4U chassis.

The license line item nobody budgets

NVIDIA sells vGPU software per concurrent user (CCU), not per GPU. Current published pricing: vApps $10, vPC $50, and RTX vWS $250 per CCU per year on annual subscription (4-year and 5-year terms discount to $43.75/$40 and $206.25/$180). Perpetual licenses are $20/$100/$450 with $5/$25/$100 annual SUMS. The license often outlives the GPU's depreciation — a 40-engineer pilot with 20 concurrent vWS seats is $5,000/year before a single server is bought, versus $1,000/year if vPC covers them.

On the client side, thin clients run $150–500 per seat and a 1080p desktop streams in roughly 1.5–3 Mbps with H.264 — meaning a 1 Gbps uplink comfortably serves 200+ concurrent sessions and the network is rarely the constraint. The constraint is always the same: license cost per seat versus GPU density, which is why the vPC/vWS split deserves a written policy before the PO.