AI server for creative agencies: image and video generation, text and GPU rendering

What one AI server does for a creative agency

A creative agency can run image generation, video generation, text models and GPU rendering on one server with NVIDIA RTX PRO 6000 Blackwell Server Edition cards, because each of these jobs fits the 96 GB of a single card. Four cards are the size we would look at first: by day one card serves the image queues in four MIG instances (isolated 24 GB slices of the card), one runs a language model and two generate video clips; at night all four render. The licence of each model, as its model card states it, decides which ones the agency may use for client work, and that check comes before the hardware.

This guide is written for the managing director and the studio's IT lead. It explains what the work asks of the hardware, which models the cards allow, and how two, four or eight cards divide the work. Our detailed guides on FLUX image generation, video generation and the GPU render server carry the full sizing for each job; here they come together on one machine.

The work in plain words

In a typical agency, four teams would use the server. Designers generate image drafts and variants for a pitch or a campaign. A motion team turns stills or prompts into short clips. Copywriters and account managers draft texts, headlines and summaries with a language model. The 3D team renders product shots and animations in Blender, V-Ray, Redshift or Octane.

These jobs differ in one way that matters for the server. An image takes seconds, a language model answers in seconds too, but a video clip or a rendered animation occupies a whole card for minutes or hours. The number of people waiting at the same moment, and the long jobs that block a card, set the card count more than the daily volume. A designer who submits ten variants while a clip renders on the same card waits for the clip.

What each job asks of a card

Hugging Face model cards of black-forest-labs and Wan-AI, read 24 October 2026; NVIDIA AI inference performance page as quoted in our FLUX guide; Blender Open Data (Blender 4.5.0, OptiX) and Chaos as quoted in our render server guide; conversations by our users-per-card rule.

Images. FLUX.1 [schnell] and FLUX.1 [dev] have 12 billion parameters and a full checkpoint of about 23 GB, and FLUX.2 [klein] 4B runs in "as little as 13GB VRAM", its card says. NVIDIA lists 0.20 images per second per Server Edition card in FP4 at batch 1, about 720 an hour by our arithmetic, without naming the FLUX variant, the resolution or the step count. As a worked example from our FLUX guide, a team that needs 1,000 images in its busiest hour needs two cards at that rate.

Video. The Wan2.2-T2V-A14B card, a mixture-of-experts model with 27B parameters and 14B active per step, says its single-GPU 720p command "can run on a GPU with at least 80GB VRAM". A 96 GB RTX PRO 6000 Server Edition meets that, and the clip then occupies the whole card. Our video guide found no published time per clip for Wan 2.2 on an RTX PRO card, so measure one clip at your settings before you size for a clip volume.

Text. gpt-oss-120b, whose checkpoint of 65.3 GB fits one 96 GB card, holds 19 conversations of 32K tokens at once on one card by our users-per-card rule: 83.04 GiB usable, less 60.8 GiB of weights, divided by 1.125 GiB of cache per conversation. With the example values of our mid-sized company guide, about 4 requests in flight per 100 employees at the busiest moment, one card covers an agency of several hundred people.

Rendering. GPU renderers split the samples of a frame across all cards, but each card holds its own copy of the scene, so 96 GB is the scene budget per node. Chaos reports V-Ray GPU scaling of 97 per cent on average with four GPUs and 91 per cent with eight, measured on RTX 3090 Ti cards, and officially supports a maximum of four GPUs per machine.

Model licences as the cards state them

For client work, every model needs a licence that covers commercial use. The table quotes what each card says, read on 24 October 2026; licence terms change, so read the card again before a production decision.

Hugging Face model cards of black-forest-labs, Qwen, stabilityai, Wan-AI and tencent and the LICENSE file of tencent/HunyuanVideo-1.5, read 24 October 2026; FLUX.2 [dev] and gpt-oss as quoted from their cards in our FLUX guide and our guide to open LLMs for European languages.

Three points stand out. The FLUX.1 Kontext [dev] card places the model under a non-commercial licence and at the same time says that generated outputs can be used for commercial purposes; using a non-commercial model in a paid workflow is a different question from owning its outputs, and your legal department answers it. For its non-commercial models such as FLUX.2 [dev], Black Forest Labs offers self-hosted commercial licences, as our FLUX guide explains. The HunyuanVideo 1.5 card names its licence only in the metadata; the licence file in the same repository, released on 21 November 2025, states that it "DOES NOT APPLY IN THE EUROPEAN UNION, UNITED KINGDOM AND SOUTH KOREA", and our video generation guide quotes its territory clause.

The licences of the renderers are a separate line. Chaos asks for one render node licence for each additional V-Ray render machine, OTOY includes 10 or more render node licences in every Octane Studio+ subscription, and Maxon includes Redshift with every Cinema 4D subscription. Our render server guide has the sources.

Since 2 August 2026, Article 50 of the AI Act has required providers of AI systems that generate synthetic audio, image, video or text content to mark the output in a machine-readable format, with exceptions for assistive editing, and deployers to disclose deep fakes. For generative systems placed on the market before 2 August 2026, the AI Act as amended by Regulation (EU) 2026/1744 sets 2 December 2026 for that marking (Article 111(4)). An agency that assembles its own generation pipeline can be its provider. Our article on the EU AI Act gives general information; national practice may differ, and your legal department decides how client contracts handle it.

Rendering and AI on the same cards

The server works when each card belongs to one job at a time. A render job uses every card it is given, a video clip takes one whole card, and an image queue needs no more than a MIG instance. If the render manager and the AI queues both claim the same card, both slow down and the larger job may run out of memory.

A plan that suits most agencies by our reading gives the cards fixed roles in working hours and hands all of them to the render manager overnight and at weekends. MIG splits a card into up to four isolated 24 GB instances, so four designers or four workflows each get a queue of their own without waiting for each other; that card stays partitioned, and the render and video jobs use whole cards. Our guide to RTX PRO 6000 MIG profiles lists the instance sizes.

The rest of the server follows from the renderer as much as from the AI models. Chaos recommends at least six physical cores per GPU and system memory of at least twice the GPU memory, which for four cards means 24 cores and 768 GB, the same figures as NVIDIA's guide for certified systems. Video offloading needs system memory as well: with one Wan 2.2 A14B job per card, a four-card server holds four sets of offloaded weights at once, as our video guide notes. The Unreal Engine path tracer is the exception among renderers, because Epic supports several GPUs only through SLI with NVLink bridges, which the RTX PRO 6000 does not have; it renders one job per card.

We build servers for agencies that combine these jobs; send us the renderers, models and the number of people who work at the same time, and our team replies with a configuration that gives each job its cards.

Configuration options by card count

Reference configurations on our AI servers page; Max-Q workstation power from our four-card Max-Q workstation guide; server power by our estimate (cards at 600 W, two AMD EPYC 9135 for two cards or 9355 for four, plus 15 per cent for the rest of the server).

A workstation with four Max-Q cards. NVIDIA states up to four RTX PRO 6000 Max-Q in one workstation, 300 W blower cards with 96 GB each, so a studio without a server room gets 384 GB of GPU memory next to the artists. Our guide to the four-card Max-Q workstation covers the towers, the power supplies and the roughly 2 kW of heat. The Max-Q is the workstation card; in a rack it belongs only in a chassis whose maker lists it, and for continuous operation we recommend the Server Edition.

Two cards. The entry server, two RTX PRO 6000 Server Edition with two AMD EPYC processors, 384 GB of DDR5 ECC and two NVMe drives, suits an agency that mostly generates images and text. One card runs four image queues in MIG instances; the other alternates between a language model, a video clip and the render jobs. Video and rendering then share one card, which limits clip volume.

Four cards. The inference node with 768 GB and four NVMe drives is the size where every job has its own card in working hours: four image queues on the first card, gpt-oss-120b on the second, one Wan 2.2 clip on each of the last two. At night, the four cards render as one node, the largest V-Ray configuration Chaos supports officially.

Two four-card servers. A larger studio that renders all day keeps the two apart: one server for the AI jobs, one for rendering, and both render at night. For V-Ray, two four-card nodes also stay within the official support of four GPUs per machine, which one eight-card server would exceed. An eight-card server suits the case where one Wan 2.2 clip must be fast: Wan's repository runs one clip across eight GPUs with FSDP and DeepSpeed Ulysses, though the RTX PRO 6000 has no NVLink and no scaling figures for PCIe-only cards are published, so measure the gain first.

Notes for the IT team

1. Queues. ComfyUI, under GPL-3.0, supports FLUX, SD 3.5 and Qwen Image and describes asynchronous queueing; run one instance per card or per MIG instance, so that a long job does not hold up the others.
2. MIG. Up to four 24 GB instances on every RTX PRO 6000 edition. NVIDIA's prerequisites are Linux, driver 575.51.03 or later, a current vBIOS and the DisplayModeSelector set to compute mode, which turns the display outputs off.
3. Remote artists. The Server Edition is the only RTX PRO 6000 edition with vGPU, from vGPU 19.0. NVIDIA RTX Virtual Workstation is licensed per concurrent user, and NVIDIA's sizing guide places two to four heavy 3D users on one card, as our vGPU profiles guide summarises.
4. Storage. Checkpoints add up: LTX-2.5's downloads alone total about 66 GiB, so a server that keeps several video and image models ready holds several hundred GB on NVMe by our estimate. Store the prompt, seed and model revision with every output, so a result can be reproduced and its licence traced.
5. Software on delivery. The server arrives on request with the operating system, the NVIDIA driver, CUDA and a container runtime. Keep the NVIDIA Container Toolkit at 1.17.8 or later and the driver at a release NVIDIA lists as fixed in its bulletin of 30 September 2026.

Power, noise and room

By our estimate, the two-card server draws about 1.8 kW at the wall with two EPYC 9135 processors and fits one 230 V 16 A feed, which carries 2.94 kW at 80 per cent. The four-card server with two EPYC 9355 draws about 3.4 kW with the cards at 600 W, or about 2.7 kW at the 450 W cap, and needs 32 A or three-phase; a capped card renders more slowly, so state the power level when you compare quotes. Both are data-centre machines in noise and heat and belong in a server room, as our guide to a GPU server in an office explains. The Max-Q workstation is the option for a studio floor.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We build servers and render nodes to order with two, four or eight RTX PRO 6000 Server Edition cards and supply four-card Max-Q workstations, assembled and tested before delivery, delivered anywhere in the EU on one EU contract and invoice, with manufacturer warranty on every component handled through us; quantities and lead time are confirmed in writing with the quote. vGPU licences come on the same invoice as the hardware. Our Private AI/ML service, with engineering by our partner Vixen.UNO, deploys the open models on-premise with controlled access. Our guides for e-commerce and media companies cover neighbouring workloads. Describe your studio's jobs and the size of your heaviest scene, and we size the cards around them.

FAQ