AI server for game studios: art, voices, lip sync and tests on RTX PRO 6000 Server Edition

What an AI server does for a game studio

A game studio uses an AI server for the production side of a game: concept art and textures, rough 3D props, placeholder voices and lip sync, help with code and localisation, and agents that play test builds. A server with four NVIDIA RTX PRO 6000 Blackwell Server Edition cards is the size we would look at first for a studio with several teams: each job fits on one 96 GB card or a share of one, so art, audio, code and testing run at the same time. Which models the studio may use in the EU is decided by each model's own licence page, and that check comes before the hardware.

This guide is for studio heads, producers and technical directors, and for the IT or pipeline lead who will run the machine. It covers the work the server takes on, the models and what their licences say, how two, four or eight cards divide the jobs, and the technical details. The server serves the team that makes the game; features inside the game that call a model while people play are a different design question and are not covered here. Our guides to an AI server for creative agencies and a GPU render server cover image, video and rendering in more depth.

A studio's AI work, job by job

Concept artists generate variations of a character, a prop or a mood piece, then paint over the one they keep. Environment artists turn reference photos into texture drafts. A 3D artist who needs fifty background props for a level can start from rough meshes made from images and spend the time on the props that the camera sees up close.

Audio and animation teams use speech synthesis for scratch dialogue, so that a scene can be timed and reviewed before the actors record, and lip-sync models that drive a face from an audio track. Programmers use a code assistant on the studio's own engine code, which for most studios must stay inside the company. Producers and the localisation team draft translations of strings and dialogue. QA runs agents that play builds again and again to find places where the character gets stuck or a level cannot be finished.

The jobs differ in one way that matters for the server. An image or a voice line takes seconds, a code answer takes seconds, but training a test agent or a batch of overnight assets occupies a card for hours. The number of people waiting at the same moment, and the long jobs that hold a card, set the card count more than the daily volume.

What each job needs on a card

Hugging Face cards of black-forest-labs, stabilityai and nvidia (Audio2Face-3D v3.0), NVIDIA TTS NIM support matrix 26.05.0 as quoted in our voice agent guide, the gpt-oss-120b repository and the ML-Agents repository on GitHub, read on 24 October 2026.

Images and textures. FLUX.2 [klein] 4B runs in "as little as 13GB VRAM", its card says, and is under Apache 2.0 with weights "available for commercial use". NVIDIA lists one RTX PRO 6000 Server Edition at 0.20 FLUX images per second in FP4 at batch 1, about 720 an hour by our arithmetic, without naming the variant, the resolution or the step count, as our FLUX guide explains. One card split into MIG instances gives each artist or tool a queue of its own.

3D props. Stable Fast 3D from Stability AI takes one image of an object, 512 × 512 pixels, and returns a textured mesh with UV unwrapping, a "relatively low polygon count" and material parameters such as roughness and metallic values. Its card says it was "trained to create a 3D model from a single image in under one second" and lists 1 billion parameters. It states no GPU memory; by our arithmetic 1 billion parameters in FP32 take about 4 GB, a small share of a 96 GB card.

Voices. Magpie TTS Multilingual is NVIDIA's speech model for 12 languages, among them English, Spanish, French, German, Italian and Japanese, under the NVIDIA Open Model License. NVIDIA's TTS NIM container reserves 10.87 GB at batch 8 and 31.55 GB at batch 32, as our voice agent guide quotes from the support matrix. Chatterbox Multilingual, 0.5 billion parameters and 23 languages under the MIT licence, adds Polish, Dutch, Swedish and others, and its card says every generated audio file carries Resemble AI's Perth watermark.

Lip sync. Audio2Face-3D v3.0 generates "3D facial animations from audio inputs": it takes audio resampled to 16 kHz and returns motion for skin, tongue, jaw and eyeballs. Its card names the Audio2Face SDK and TensorRT as the runtime, Linux and Windows, and lists the Blackwell architecture among those supported, with tested GPUs that include the L40S and a "Pro 6000 Blackwell" without a named edition. It warns that the model "may generate inaccurate lip poses given low-quality audio input", so animators review the result.

Code and text. gpt-oss-120b, with a checkpoint of 65.3 GB, fits one 96 GB card and holds 19 conversations of 32K tokens at once by our users-per-card rule: 0.9 × 95.6 GiB driver-visible memory − 3 GiB per card = 83.04 GiB usable, less 60.8 GiB of weights, divided by 1.125 GiB per conversation. For coding agents that work on large parts of a code base, our server for coding agents puts a pair of cards with Qwen3-Coder-Next at no more than about 35 sessions of 64K tokens at 50 tokens per second each, an upper bound by our estimate for this mixture-of-experts model. Translations of strings and dialogue are covered in our guide to machine translation on-premise.

Test agents. Unity's ML-Agents toolkit "enables games and simulations to serve as environments for training intelligent agents" and names "automated testing of game builds and evaluating different game design decisions pre-release" among its uses. Its algorithm implementations are "based on PyTorch", and its repository is under Apache 2.0, with release 23 of 28 August 2025 as the latest. The repository gives no GPU figure, so a first agent on one card shows the size.

What the licence pages say

For a commercial game, every model needs a licence that covers commercial use where the studio works. The table quotes the licence pages and model cards as read on 24 October 2026; terms change, so read them again before a production decision. We state only what the pages say, and your legal department decides on each use.

Hugging Face model cards and licence files of black-forest-labs, stabilityai, tencent (LICENSE, release date 13 June 2025), ResembleAI and nvidia, and NVIDIA's Magpie TTS blog of 10 August 2026, read on 24 October 2026; FLUX.2 [dev] and gpt-oss as quoted in our FLUX guide and creative agency guide.

Two entries need a closer look. The Hunyuan3D 2.1 licence, released on 13 June 2025, defines its "Territory" as "the worldwide territory, excluding the territory of the European Union, United Kingdom and South Korea", and its preamble states in capitals that it does not apply there. By that wording, the licence does not cover a studio that uses the model in the EU. The same licence also requires a licensee whose products or services had more than 1 million monthly active users in the calendar month before the release date to request a licence from Tencent, and forbids using the model or its outputs to improve other AI models.

Stable Fast 3D is free "for research, non-commercial, and commercial use" for organisations with annual revenue of USD 1,000,000 or less; above that, its card asks for an enterprise licence from Stability AI before commercial use, including for fine-tuned versions. A studio that grows past that threshold during a project needs the enterprise licence from that point, so put the revenue check into the project plan.

Voices, actors and the AI Act

Synthetic speech in a studio usually means scratch lines that are replaced by recorded actors, or voices for minor characters. Recordings of an identifiable actor are personal data under the GDPR, and the actor's contract sets what the studio may record, train on or synthesise; that question belongs to your legal department before any voice model touches a recording. Among NVIDIA's speech models, Magpie TTS Zeroshot is listed as access-restricted in NVIDIA's support matrix, as our voice agent guide notes.

Article 50 of the AI Act has applied since 2 August 2026. Providers of AI systems that generate synthetic audio, image, video or text mark the output in a machine-readable format, with exceptions for assistive editing, and deployers 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)). A studio 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 it applies to the studio's tools and to the game.

We build servers for studios that combine these jobs; send us the tools your teams use, the number of people who work at the same time and the models you plan, and our team replies with a configuration that gives each job its card.

Two, four or eight cards for a studio

Reference configurations on our AI servers page; Max-Q workstation from our four-card Max-Q workstation guide; server power by our estimate (cards at 600 W, two AMD EPYC 9135, 9355 or 9555 for two, four or eight cards, plus 15 per cent for the rest of the server).

A four-card workstation. NVIDIA states up to four RTX PRO 6000 Max-Q in one workstation, 300 W blower cards with 96 GB each. A studio without a server room gets 384 GB of GPU memory under a desk, and our guide to the four-card Max-Q workstation covers the towers 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 in a rack we recommend the Server Edition.

Two cards. Our entry server has two RTX PRO 6000 Server Edition cards with two AMD EPYC processors, 384 GB of DDR5 ECC memory and two NVMe drives. One card runs image and voice queues in up to four MIG instances of 24 GB; the other carries gpt-oss-120b for code and text by day and trains test agents at night. It fits one 230 V 16 A feed.

Four cards. The inference node, with 768 GB of memory and four NVMe drives, gives each team its card in working hours: art, audio, code and QA. At night the art and audio cards can render or generate batches of assets, and the QA card keeps training agents. It needs a 32 A or three-phase feed at full power, or runs at the 450 W cap at about 2.7 kW.

Eight cards. An eight-card server suits a studio that runs several projects at once, or a code model that needs more memory than a pair. Our comparison of one eight-GPU server or two four-GPU servers sets out when two machines are the better choice, for example to keep one running while the other is serviced.

For the studio's IT and pipeline team

1. Queues. ComfyUI, under GPL-3.0, supports FLUX and Qwen Image and describes asynchronous queueing, as our FLUX guide notes; run one instance per card or per MIG instance, so that a long batch does not hold up an artist.
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. NVIDIA's MIG guide states that NCCL is not supported with MIG, so multi-card training uses whole cards.
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, as our vGPU profiles guide explains.
4. Engines and rendering. Epic supports several GPUs for the Unreal Engine path tracer only through SLI with NVLink bridges, which the RTX PRO 6000 does not have, so it renders one job per card, as our render server guide notes.
5. Storage. Keep models, generated assets and the prompt, seed and model revision of every output on NVMe, so that a result can be reproduced and its licence traced. Our guide to NVMe storage for a GPU server covers drive classes.
6. 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 security bulletin of 30 September 2026. NVIDIA's speech NIM containers need an NVIDIA AI Enterprise licence for production use, while open weights can also run outside NIM.

Where the server goes: power and heat

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 the eight-card server with two EPYC 9555 about 6.3 kW. These are data-centre machines in noise and heat and belong in a server room or a rented rack, as our guide to a GPU server in an office explains. The Max-Q workstation is the option for a studio floor, at about 2 kW of heat.

Ordering the server from us

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 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. NVIDIA AI Enterprise and vGPU licences come on the same invoice as the hardware. Our Private AI/ML service, built by our engineering partner Vixen.UNO, sets up the open models and the code assistant on-premise with controlled access. Describe your studio's jobs and how many people use them at once, and we size the cards around them.

FAQ