AI server for patent attorneys: drafting, prior-art search and translation before filing

What an AI server does for a patent firm

An AI server for patent attorneys runs drafting, prior-art search and translation models on hardware the firm or the corporate IP department controls, so that invention disclosures, draft claims and search queries never reach a third party before the application is filed. One server with two NVIDIA RTX PRO 6000 Blackwell Server Edition cards carries a long-context drafting model on one card and the search, translation and figure models on the other. Four cards suit a firm with many attorneys working on long specifications at the same time, or an IP department that also indexes a large patent corpus.

We wrote this for the managing partner of a patent firm or the head of an IP department who decides whether such a server is worth having, and for the IT team that runs it. The first sections explain why the confidentiality point is specific to patent work and what each workload needs; the configurations and IT notes follow. Figures are published measurements with their source, or our estimates, labelled as such. Our guide to a private LLM for law firms covers DMS search, matter access and the CCBE guide for lawyers; this page covers what is particular to patents.

Article 54 EPC, the epi guidelines and confidentiality

Article 54(1) of the European Patent Convention treats an invention as new "if it does not form part of the state of the art", and Article 54(2) holds the state of the art "to comprise everything made available to the public by means of a written or oral description, by use, or in any other way, before the date of filing of the European patent application." Article 55(1) disregards a disclosure only in narrow cases, such as one that occurred no earlier than six months before filing and was due to "an evident abuse in relation to the applicant or his legal predecessor". Whether a disclosure to an AI provider makes anything available to the public is a legal question that depends on the provider's terms, and the firm's own legal assessment decides it. A server inside the firm keeps the question from arising.

The Institute of Professional Representatives before the European Patent Office (epi) adopted its Guidelines on the use of generative AI in the work of patent attorneys at its Council meeting C98 on 16 November 2024 and published them in epi Information 4/2024. Guideline 2a asks members, "to the extent called for by the circumstances", to "ensure adequate confidentiality of training datasets, instruction prompts and other content transmitted to AI models" and adds: "If there is doubt that confidentiality will be maintained to a level that is appropriate to the prevailing context", "the AI model in question should not be used". Guideline 2b states that "Members must inform themselves about the likelihoods and modes of non-confidential disclosures". Under Guideline 3b members "must check any work product produced using generative AI for errors and omissions", and Guideline 4 requires them to "establish, in advance of using generative AI in their cases, the wishes of their clients". Guideline 6 asks members, where the nature of confidentiality in the model warrants it, to "establish mutually independent user accounts for the work of respective clients". The epi Professional Conduct Committee added in epi Information 3/2025 that "a generalised agreement over the use of AI tools is not sufficient to satisfy the requirement of Guideline 4 unless the terms identify the areas of work in which AI tools are to be used", and states that its opinion is not binding on the disciplinary bodies.

A model on the firm's own server answers the transmission question for prompts and documents, because nothing leaves the network unless someone enables an outside service. It does not answer the client-consent and checking duties, which stay with the attorney. This is general information as of October 2026 from the EPC text on epo.org and the epi pages; national rules and your professional body's guidance may add duties, and your own legal assessment decides.

Four workloads and the GPU each needs

Model facts from the Hugging Face model cards (PaECTER read 24 October 2026; the others as in our guides on law firms, embeddings, translation and document AI); GPU choice is our reading of the memory each model needs.

Drafting from an invention disclosure

A drafting session carries a lot of text: the inventor's disclosure, the closest prior-art documents and the firm's own claim templates, and the model writes claims, a description and a summary from them. That makes context length the memory driver. gpt-oss-120b, released under Apache 2.0, accepts up to 131,072 tokens, and by the figures in our law-firm guide its 16-bit KV cache takes 1.125 GiB for a 32K conversation and 4.5 GiB at 128K. By our users-per-card rule, one RTX PRO 6000 leaves 22.2 GiB of cache beside the 60.8 GiB of weights, enough for 19 sessions at 32K or four at 128K. A pair of cards sharing one copy of the model keeps 105.3 GiB, about 23 sessions at 128K.

Head count matters less than it seems. With the example values of our guide to private ChatGPT server sizing by company size, about 4 per cent of staff have a request in flight at the peak, so a firm of 100 people has about four, which one card holds even at 128K. A firm whose attorneys work with full specifications and several cited documents in every request should plan the pair. Our guide to long-context LLM hardware compares the cache per token of other models.

The model drafts; the attorney decides. Every claim set and description goes through the same review as a junior's draft, as Guideline 3b requires, and the server keeps the prompt, the sources and the answer in a log so that the review can see what the model was given. Long outputs take time to generate, so a full description is a job the attorney starts and returns to; the response time on your own disclosures is what a pilot measures.

We supply the RTX PRO 6000 Server Edition and build servers with two, four or eight of them; send us the number of attorneys and the length of your typical specifications, and our team replies with a configuration.

Semantic prior-art search over a patent corpus

Keyword and classification searches stay in the official databases. What a firm adds on its own server is semantic search: the attorney pastes a claim or a paragraph of the disclosure and receives the documents that say the same thing in other words. The query is itself a description of the unfiled invention, which is why it belongs on the same server as the drafting model. The corpus can be the firm's own files and the patent documents it already holds or licenses.

Two kinds of embedding model fit. PaECTER, published by the Max Planck Institute for Innovation and Competition on Hugging Face, is a patent similarity model built on Google's BERT for Patents, with 0.3B parameters, 1,024-dimension vectors and an input of 512 tokens under Apache 2.0; its card names prior-art search among its uses and tags the model as English. Qwen3-Embedding-0.6B, also Apache 2.0, supports more than 100 languages and 32K tokens of input, as our guide to embedding and reranker servers lists. A reranker then orders the top results.

Indexing is the larger job. Take 2 million documents in 10 chunks of 512 tokens each as an example, 20 million chunks. NVIDIA publishes 144.3 chunks of 512 tokens per second on an L40S for its 1B embedding model in FP8, so by our arithmetic the first run takes about 38.5 hours on that card; we found no published figure for the RTX PRO 6000 or for PaECTER, so time a sample on the target card. At 1,024 dimensions in 32-bit, each vector takes 4 KiB and the 20 million about 82 GB before index structures and metadata, held in server RAM unless the vector database keeps it on disk, as our vector database comparison explains. Afterwards only new publications are embedded, a small weekly load.

Translation for unitary effect and validation

Translation is a steady part of patent work. Council Regulation (EU) No 1260/2012 on the translation arrangements for unitary patent protection requires, during its transitional period, that a request for unitary effect include "a full translation of the specification" into English where the language of the proceedings is French or German, or into another official EU language where it is English. Article 6(2) states that "The text of such translations shall have no legal effect and shall be for information purposes only", and Article 6(5) lets the transitional period lapse at the latest 12 years after the Regulation's date of application. National validations of European patents have their own translation rules, which this page does not cover.

A translation model on the server produces the draft, and a translator or attorney revises it. The licence decides the shortlist: as our guide to on-premise machine translation quotes, NLLB-200 is a non-commercial research model, while EuroLLM covers all 24 official EU languages under Apache 2.0, and TranslateGemma comes under Google's Gemma Terms of Use. That guide sizes one 48 GB card for about 1 million words a day with a 9B to 12B translation LLM in 16-bit, by its estimate of two output tokens per word. A 48 GB MIG instance of an RTX PRO 6000 holds such a model by the same memory figures; its speed in a MIG instance is not published. Test the terminology of your technical fields, since patent language differs from the texts these models were evaluated on.

Figures and drawings

Drawings need a vision-language model: to describe a figure for the brief description of the drawings, to check that every reference sign in the text appears in a figure, or to read the figures of a cited document. Qwen3.8-27B, under Apache 2.0, takes 30.9 GB in FP8 and by our document AI guide holds about 150 pages in flight on one RTX PRO 6000. This is occasional work, so it can run at night on the drafting card, or take the place of the translation model on the second card when needed. Our guide to vision-language models for document AI gives the image tokens per page and the throughput method.

Configuration options by card count

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 twice the GPU memory.

Reference configurations on our AI servers page; sessions by our users-per-card rule with gpt-oss-120b in a 16-bit cache; power by our estimate (cards at 600 W, processors, plus 15 per cent for the rest of the server).

Two cards. The entry server has two RTX PRO 6000 Server Edition, two AMD EPYC processors, 384 GB of DDR5 ECC and two NVMe drives. One card runs gpt-oss-120b for drafting. The other is split by MIG into two isolated 48 GB instances, one for the translation model and one for the embedding model and reranker, and the initial indexing of a large corpus runs at night. It suits a firm of up to about 100 people at the example values above.

Four cards. The inference node with 768 GB and four NVMe drives gives drafting a pair, translation and figures a card, and search and indexing a card of their own. An IP department with a large corpus and many long specifications would start here.

Two servers. If drafting is part of daily work for many attorneys, two four-card servers that each carry the peak let one fail or be updated while the other answers, as our guide to LLM high availability with two GPU nodes describes.

The RTX PRO 6000 Max-Q is the workstation card, a 300 W blower card; in a rack it belongs only in a chassis whose maker lists it, and for continuous operation we recommend the Server Edition. An attorney who wants a model at the desk can use a workstation with one Max-Q, but the firm then has one more place where disclosures are stored. We deploy the models, search and the per-client separation through our Private AI/ML service; tell us which workload you would start with, and we propose the pilot and the server for it.

Notes for the IT team

1. Per-client separation. Give each client, or each matter, its own index partition, API key and log, so that a search for one client never retrieves another client's unfiled documents. This is the server-side counterpart of Guideline 6.
2. No outside calls. Public model APIs stay off, and any external call needs an explicit decision that is visible in the log. An LLM gateway with keys and a query log in front of the engines records who sent what.
3. Retention. Prompts and drafts hold the invention, so set how long the logs are kept and who may read them before go-live, and include test copies.
4. Network and versions. 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 behind the gateway. Run vLLM 0.31.0 or later, 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.
5. Software on delivery. The server arrives on request with the operating system, the NVIDIA driver, CUDA and a container runtime installed; our guide to securing a GPU server covers the rest of the hardening.
6. Licences of the models. Check each model's licence file before use; our comparison of open LLM licences for company use quotes the clauses that matter.

Power 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 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 rack server of this kind belongs 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. 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; whether that suits your clients is part of the confidentiality assessment above.

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, four or eight 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. Our Private AI/ML service, with engineering by our partner Vixen.UNO, deploys the drafting, search and translation models with per-client separation. The RTX PRO 6000 page lists the editions.

FAQ