The Drupal AI module stack ships with provider plugins for OpenAI, Anthropic, Azure OpenAI and “OpenAI-compatible API”. The last option is what lets you redirect every call to a vLLM endpoint on the RTX 5060 Ti 16GB hosted on our UK dedicated GPU hosting, turning per-token charges into a fixed monthly fee. Drupal dominates UK public sector and higher-education builds where data residency is mandatory, and a single Blackwell card with 4608 CUDA cores, 16 GB GDDR7 and native FP8 is enough to serve an entire estate of content editors.
Contents
- AI module stack
- Provider configuration
- Throughput on editorial workloads
- Cost vs SaaS providers
- Public sector fit
AI module stack
| Module | Purpose | Recommended model |
|---|---|---|
| ai | Core provider abstraction | n/a |
| ai_content_suggestions | Draft body copy, summaries, titles | Llama 3.1 8B FP8 |
| ai_translate | Content translation between languages | Qwen 2.5 14B AWQ |
| ai_image_alt_text | Alt text from uploaded images | LLaVA 1.6 7B |
| ai_assistant | Back-office admin helper | Mistral 7B FP8 |
| ai_search | Semantic site search | BGE-M3 embedder |
Provider configuration
Under Configuration -> AI -> Providers, add an OpenAI-compatible provider with base URL https://llm.example.ac.uk/v1, a static key and your served model name. Drupal then lists it in every module’s provider dropdown. No patches required; the integration lives entirely in module configuration.
Throughput
| Editorial task | Input | Output | Time on 5060 Ti |
|---|---|---|---|
| 600-word page draft | 200-word brief | 900 tokens | 8.0 s |
| Image alt text | 1 image | 40 tokens | 0.6 s |
| EN to CY translation | 600-word article | 700 tokens | 10 s (Qwen 14B) |
| Metadata (title + description) | Article body | 80 tokens | 0.3 s |
| Admin Q&A against site content | RAG prompt | 300 tokens | 2.7 s |
With 16 concurrent editors on Llama 3.1 8B FP8 the card sustains around 720 t/s aggregate, which covers even large councils or universities during peak publication windows.
Cost vs SaaS
| Workload / month | OpenAI GPT-4o | Azure OpenAI (UK South) | Self-hosted 5060 Ti |
|---|---|---|---|
| 2k page drafts, 10k alt-texts, 5k translations | ~£450 | ~£500 | Flat £300 |
| 20k metadata generations | ~£60 | ~£70 | Same box |
| Annual | ~£6,100 | ~£6,800 | £3,600 |
Public sector fit
UK central government, NHS Digital and many Russell Group universities run Drupal. A dedicated 5060 Ti in a UK data centre ticks the data sovereignty box, satisfies the single sub-processor requirement in most public sector DPAs, and removes the “model training on my data” concern that still blocks many procurement teams from adopting hosted AI services. GDPR Article 28 contracts are simpler when there is no US transfer clause to negotiate.
UK-sovereign Drupal AI backend
Blackwell 16GB for the Drupal AI module suite. UK dedicated hosting.
Order the RTX 5060 Ti 16GBSee also: AI-powered CMS, internal tooling, FP8 Llama deployment, document Q&A, RAG stack install.
