On demand
£6.95
per hour plus VAT
When the machine is not reserved.
For Architects
Run a large AI model on your practice's documents - without any of them leaving the UK
For RIBA chartered practices of 50-plus staff, multidisciplinary consultancies with architecture arms, and architecture teams inside contractors and developers.
This is not a render farm. MetalCloud does not run Enscape, Lumion, V-Ray, Twinmotion or any render engine. It does not read or write model geometry, IFC or RVT. It is not a Revit or ArchiCAD plugin. It is a model endpoint reached over an API.
01
The problem
Specifications, Approved Documents, planning conditions, JCT contracts and consultant reports run to thousands of pages per scheme. None of it can be pasted into a public AI service where client confidentiality clauses or professional indemnity wording say otherwise. Practices are stuck choosing between not using AI and using it in a way they cannot defend at renewal. A Building Safety Act gateway, a competition under NDA, or a PI proposal that asks about AI use brings the same documents into the same corner.
02
What you can actually run
What you can actually run is a 70B-class open-weight model held in memory at full precision, pointed at the practice's own specifications, Approved Documents, consultant reports and contracts. The answers can be drawn from the practice's own precedent, not from a generic model's training data. Copilot is not a substitute for that: it processes off-premises and does not hold your document set in memory at full precision on UK-hosted, single-tenant hardware. If the model is wrong, a human remains in the loop. That is always the position.
03
Named workflows
- Specification against drawing-package consistency Read the written specification as a document set. Geometry is out of scope.
- Building Regulations and Approved Document questions Answered against the practice's own precedent and the documents you supply.
- Planning condition discharge tracking Keep conditions and evidence together as long files on one machine.
- Fee proposal and bid drafting Draft from the practice's own past submissions.
- Consultant report interrogation Ground investigations, surveys and fire strategies, read as documents.
- O&M manual and handover assembly Assemble handover text from material the practice already holds.
04
Why this hardware
This is a capacity machine: 512GB of unified memory on a single tenant. The CPU and GPU share one pool, so a large model and a large document set can sit in memory together at full precision. A 70B-class open-weight model is the named size class this configuration is used for.
It is not sold as a high-throughput service. Apple Silicon does not suit many simultaneous users, and this page makes no claim about tokens per second, response time, or concurrency. The fit is single-tenant, privacy-led, batch-friendly work: overnight processing, long documents, and a model that stays on hardware only your team can reach.
Dedicated monthly use is one customer on one machine. That is the product.
05
Buying one is not currently an option
Apple launched the M3 Ultra Mac Studio in March 2025 advertising up to 512GB of unified memory, and marketed it as able to run models with over 600 billion parameters entirely in memory.
In March 2026, MacRumors, 9to5Mac and Ars Technica reported that Apple withdrew the 512GB configuration. In May 2026, Apple Insider and others reported that the 256GB configuration was withdrawn as well. The M3 Ultra Mac Studio is now sold in a single 96GB configuration.
Secondary-market listings for 512GB units have been reported around $25,000, roughly £19,000-20,000, where they appear at all. Apple has given no public reason and has not said whether the change is permanent. The M5 Ultra Mac Studio has not shipped.
MetalCloud did not originate these facts. They are a matter of public reporting. This page will be rewritten if Apple restores a high-memory configuration.
Currently one 512GB machine. A dedicated monthly customer occupies it in full. Allocation is first-committed.
Correct as of 22 August 2026. Sourced from public reporting, not from MetalCloud.
06
Data handling
The machine is hosted in the UK. On the reserved monthly tier it is single-tenant: your team reaches the endpoint, and another firm does not share the box for that period. Execution is hardware-isolated at the machine.
MetalCloud does not hold SOC 2. ISO 27001 is not certified. A data processing agreement is available on request. Operator access, logging and retention are set out in that agreement and on the technical call, not invented here.
When you stop, you stop. What happens to residual data on disk is a DPA and offboarding question, and we will answer it in writing before you commit.
- Where it sits
- UK-hosted Mac Studio.
- Who reaches it
- Your team, on a reserved booking. Single tenant.
- What is logged
- Set out in the DPA and on the technical call. Not invented here.
- Agreement
- DPA available on request. ISO 27001 is not certified. SOC 2 is not held.
07
Pricing
Two published prices, both plus VAT. On demand is £6.95 per hour. Reserved monthly is £2,995 per month. There is currently one 512GB machine. A reserved monthly customer occupies it in full. Hourly work is offered when the machine is not reserved. Allocation is first-committed.
Reserved
£2,995
per month plus VAT
Occupies the machine for the month.
Headline monthly cost is the wrong frame for a practice. £2,995 plus VAT reserved is the published monthly price for the machine, which is one machine. Spread across fee earners in a 50-plus practice, it is a per-person figure the managing partner can put next to PI and IT. Hourly use at £6.95 plus VAT exists when the machine is not reserved. Smaller practices will not want this number, and this page is not aimed at them.
Currently one machine, allocated on a first-committed basis. Neither price includes VAT.
08
What MetalCloud is not
If you typed GPU into a search bar looking for pictures, stop here.
- Not a rendering or visualisation service.
- Does not run Enscape, Lumion, V-Ray, Twinmotion or any render engine.
- Does not read or write model geometry, IFC or RVT.
- Not a Revit or ArchiCAD plugin.
09
Questions architects teams ask
Yes, as documents you supply. It does not read the drawing model.
On a UK-hosted Mac Studio. Reserved use is single-tenant.
No. It is an API endpoint, not a plugin, and it does not read RVT or IFC.
That client material is processed on UK-hosted, single-tenant hardware, that a human remains in the loop, and that a DPA is available. The insurer will ask for that in writing.
No.
Offboarding is written into the DPA and confirmed before you commit.
Book a 20-minute technical call
If a gateway, an NDA competition or a PI renewal is the reason you are here, book a 20-minute technical call.