AI server for construction and architecture: tender documents, BIM and reality capture

What an AI server does for a construction company or architecture practice

An AI server for a construction company or an architecture practice runs three kinds of GPU work on hardware the firm controls: a language model that reads tender packages, bills of quantities and contracts, the reconstruction of sites and buildings from drone photos and laser scans, and image generation for design variants. A server with two NVIDIA RTX PRO 6000 Blackwell Server Edition cards carries a tender assistant and the reconstruction jobs of one survey team; four cards add a pair of cards that shares one copy of the model for long documents, and two cards for reality capture. Project documents, site photos and unpublished designs stay on the firm's own server.

We wrote this for the managing director or the head of estimating, surveying or design who has to judge what such a server does, and for the IT team that runs it. Virtual CAD workstations on vGPU are a separate topic, covered in our guide to a GPU platform for engineering firms; final-frame rendering is covered in our GPU render server guide.

Three workloads with different limits

Our summary; document capacity by our users-per-card rule; COLMAP FAQ and gsplat README, read 24 October 2026; FLUX figures as in our FLUX guide.

The three rarely compete for the same hours: estimators read tenders in office hours, reconstruction jobs can wait for the night, and image generation comes in brief bursts during design reviews.

Tender packages and bills of quantities

A tender package for a building project combines the bill of quantities, technical specifications, contract conditions, drawings lists and the client's questions and answers. An estimator has to find the requirements that change the price or the risk: unusual specifications, penalties, deadlines, required certificates and the positions that depend on each other. A local language model can read the whole package, list these points with the page where they appear, draft questions for the client and compare subcontractor offers against the bill of quantities. A person checks every result before it enters a bid.

Document length decides the size of the server. gpt-oss-120b, published under Apache 2.0 with 117B parameters of which 5.1B are active, has a maximum context of 131,072 tokens in its configuration, and its checkpoint files add up to about 65 GB, 60.8 GiB. By our users-per-card rule, which takes 90 per cent of the 95.6 GiB the driver reports, less 3 GiB per card and less the weights, one RTX PRO 6000 Server Edition keeps 22.2 GiB for the KV cache, the working memory the model fills for every document it has open. A document at the full context takes 4.5 GiB of 16-bit cache, so one card holds four such documents in flight. Two cards that split one copy of the model between them (tensor parallelism, a tensor-parallel pair) hold 23. Shorter questions on the same server need far less, as our guide to long-context LLM hardware explains. A specification longer than the context window is read chapter by chapter with retrieval over the whole package, as in our guide to RAG on company data.

Bid deadlines bring peaks: several estimators open large packages in the same week. Reading packages overnight as a batch job, with the results waiting in the morning, takes the peak off the cards; our guide to batch LLM inference on-premise describes the queue. Tender documents in several European languages need a model tested in those languages, as our comparison of open LLMs for European languages suggests.

Public clients add a second reason to work with models of buildings. Article 22(4) of Directive 2014/24/EU on public procurement states that "for public works contracts and design contests, Member States may require the use of specific electronic tools, such as of building information electronic modelling tools or similar", and the contracting authorities then offer alternative means of access until such tools are generally available. Whether and how your country uses that option is set in national law; your legal counsel confirms it for a given tender.

BIM models and model checks

BIM models carry the quantities and properties a tender refers to. A language model does not read a model file natively, but it can work on exports: element lists with their properties, quantity take-offs and clash reports. A typical task is to compare the properties in the export with the specification and list the mismatches for the BIM coordinator. We found no published benchmark for this kind of check, so treat it as a pilot on one project before you rely on it.

We supply two- or four-card RTX PRO 6000 Server Edition servers and set up the model and retrieval through our Private AI/ML service; send us the size of a typical tender package and the number of estimators, and our team replies with a configuration.

Reality capture from drone photos and laser scans

Photogrammetry turns overlapping photos into a point cloud and a mesh; 3D Gaussian splatting turns the same photos into a scene that renders in real time. Both start by finding the camera positions, and both use the GPU heavily.

COLMAP is the open structure-from-motion and multi-view stereo pipeline, licensed under the new BSD licence; the licence page notes that building it with third-party dependencies may affect the resulting licence. Its FAQ states that the stereo step "runs on the GPU using CUDA and puts the GPU under heavy load", and that without CUDA all functions except dense reconstruction run on the CPU, much more slowly. Feature extraction, matching and dense reconstruction can run on several GPUs by listing their indices, with one thread per GPU. Two details matter for a server: the distribution packages "do not come with CUDA or HIP/ROCm support", so COLMAP is built from source or run from its pre-built Docker image with CUDA support, and a build for these cards needs a CUDA toolkit that knows Blackwell, which started with CUDA 12.8, as our guide to RTX PRO Blackwell software support explains.

GPU memory in COLMAP depends on the settings. If patch-match stereo runs out of memory, the FAQ suggests a lower maximum image size or fewer source images. For SIFT matching on the GPU it gives an estimate of 4 × n² + 4 × n × 256 bytes for n matches: by our arithmetic about 4.3 GB at 32,768 matches and about 17.2 GB at 65,536. A 96 GB card leaves room for high-resolution photos and large match limits, where a smaller card forces lower settings.

3D Gaussian splatting comes from the 2023 paper "3D Gaussian Splatting for Real-Time Radiance Field Rendering" in ACM Transactions on Graphics, with code published by Inria's GRAPHDECO team. The original code asks for a GPU of compute capability 7.0 or higher and "24 GB VRAM (to train to paper evaluation quality)", and less for shorter training. Its licence allows use "non-commercially", for research and evaluation, so a construction company cannot use it commercially without the licensors' prior consent. Its setup also assumes CUDA SDK 11, which predates Blackwell. gsplat, from the nerfstudio project, is published under Apache 2.0, describes itself as a CUDA-accelerated library for rasterising Gaussians, and reports training with up to four times less GPU memory and up to 15 per cent less time than the official implementation. Its README names build compatibility with CUDA 12.8 and 13.2; multi-GPU dense training is on its main branch and, as of 24 October 2026, not yet in a PyPI release.

COLMAP installation page, FAQ and licence page (version 4.3.0.dev0 of 11 October 2026); graphdeco-inria/gaussian-splatting README and LICENSE.md; nerfstudio-project/gsplat README; all read 24 October 2026.

By memory, a 96 GB card holds four splatting jobs at the 24 GB of the original paper. With MIG, an RTX PRO 6000 splits into two isolated 48 GB instances or four of 24 GB, each with its share of the compute, so two surveyors can train scenes at the same time without affecting each other. Survey software from commercial vendors has its own GPU and operating-system requirements; check them with the vendor before you plan a server around it.

Site photos often show people and vehicles. Under Article 4(1) of the GDPR, personal data is "any information relating to an identified or identifiable natural person", so faces and number plates in drone photos can make a data set personal data. Keeping the photos on your own server simplifies access rules and deletion; whether you need to blur, how long you keep the images and whether a data protection impact assessment is needed are questions for your data-protection officer.

Design variants and visualisation

Architects can use image models to sketch variants of a facade, a material or an outdoor area before a scene is modelled in detail. Our guide to FLUX GPU requirements gives the numbers: FLUX.1 [schnell] is under Apache 2.0, while FLUX.1 [dev] and FLUX.2 [dev] carry non-commercial licences, and NVIDIA lists one RTX PRO 6000 Server Edition at 0.20 FLUX images per second in FP4 at batch 1, about 720 an hour, without stating resolution or steps. By the same guide, a 24 GB card runs the 12-billion-parameter FLUX.1 models in FP8 or FP4, so a 24 GB MIG instance next to other work on the same card is a candidate for a design team; test it with your resolution first.

Final renders of a building follow other rules: GPU renderers split the samples of a frame across cards, but every card has to hold the whole scene, as our render server guide explains. Designers who model and review at a desk work best on a workstation with an RTX PRO 6000 Max-Q or a smaller RTX PRO card, as our guide to GPUs for CAD workstations shows. The Max-Q is the 300 W workstation edition; in a rack it belongs only in a chassis whose maker lists it, and for continuous operation we recommend the Server Edition.

Configuration options

The options follow the reference configurations on our AI servers page and NVIDIA's guide for certified systems: two processors, at least six cores per GPU and system memory of at least twice the GPU memory.

Reference configurations on our AI servers page; documents in flight by our users-per-card rule with gpt-oss-120b at 131,072 tokens and 16-bit cache; power by our estimate (cards at 600 W, two processors, plus 15 per cent for the rest of the server), as in our sibling server guides.

Two cards suit a mid-sized contractor or practice that starts with tenders: one card runs gpt-oss-120b for the estimators, the other reconstructs the scans of the day overnight and serves image generation in MIG instances during the day. Four cards suit a firm with its own survey team: the tensor-parallel pair keeps 23 long documents in flight for bid peaks, and two cards run COLMAP across both GPUs or two splatting jobs at once. Server plus workstations keeps interactive design at the desk and puts everything that runs for hours into the server room.

Notes for the IT team

1. Storage. Drone flights and laser scans produce large image sets per project, and reconstruction adds intermediate files such as depth maps. Plan fast NVMe for the active projects and a file server for the archive; our guide to NVMe storage for an AI server covers the drives. The size depends on your flights and scanners, so we size it from a sample project.
2. Builds. Build COLMAP and gsplat against a CUDA toolkit of 12.8 or later and record the versions; COLMAP's build option CMAKE_CUDA_ARCHITECTURES defaults to the GPU of the build machine.
3. Interface. vLLM serves an OpenAI-compatible API, so the tender assistant, document management and chat front end call one interface. Put an LLM gateway with keys and a query log in front of it, and keep project folders separated by access rights in the retrieval index.
4. Security. vLLM's --api-key protects only paths under /v1, /v2, /inference and /cohere, while routes such as /invocations and /metrics answer without the key (vLLM 0.32.0 security guide), so keep the engines on an internal network. Keep the NVIDIA Container Toolkit at 1.17.8 or later and the driver at a release NVIDIA's bulletin of 30 September 2026 lists as fixed.
5. Software on delivery. The server arrives on request with the operating system, the NVIDIA driver, CUDA and a container runtime installed.
6. AI Act. For tender reading and image generation, Article 4 of the AI Act, as amended by Regulation (EU) 2026/1744, requires providers and deployers to take measures to support the AI literacy of their staff, as our guide to AI literacy training under Article 4 explains, and Article 50 adds transparency duties for some generated content. Our article on the EU AI Act gives general information; national practice may differ and your legal department decides.

Power and room

By our estimate, the two-card server draws about 1.8 kW at the wall and fits one 230 V 16 A feed, which carries 2.94 kW at 80 per cent. The four-card server draws about 3.4 kW with the cards at 600 W and needs 32 A or three-phase, or about 2.7 kW with the cards capped at 450 W. A site office or a design studio is no place for these machines; they belong in a server room or a data centre, as our guide to GPU server power on 16 A, 32 A and three-phase feeds works through. A power estimate is confirmed by a measurement on the delivered system. Without a server room of your own, our Private AI/ML service runs the platform on dedicated hardware in a Tier-3 data centre in Lithuania, inside the EU.

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 the servers to order with two or four RTX PRO 6000 Server Edition cards, 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. We also supply RTX PRO 6000 Max-Q and other RTX PRO cards for design workstations. Our Private AI/ML service, delivered with our engineering partner Vixen.UNO, deploys the language model, retrieval over your project documents and the connections to your systems.

FAQ