Breaking

Zaha Hadid Architects runs in-house fine-tuned AI locally to speed design

June 26, 2026 at 16:00 EDT

Sales & Business
Media Generation
Infra & Chips

Architecture firm Zaha Hadid Architects (ZHA) is running AI models fine-tuned on more than 40 years of its own proprietary data on local GPU workstations, accelerating its design process while keeping confidential data protected, in a case study published by NVIDIA.

Zaha Hadid Architects · AI in Design

From a week to under 24 hours: ZHA builds private, on-site AI for design

The architecture firm fine-tunes its own AI models on four decades of proprietary work — run locally to keep client data off external servers — collapsing the client-feedback cycle to near real time.

The revision cycle, drawn to scale

Each block ≈ one day. The client-facing visual loop went from roughly a week of manual work to under a single day — in some cases near real time.

~7 days

Before (manual)

→

<24 hrs

After (AI loop)

40 yrs

Of proprietary portfolio feeding fine-tuned models

60M

Polygons in the largest scenes now handled

4 GPUs

Per HP Z workstation, all processing on-site

10+

Custom extensions inside ZSPACE Create

How the private stack fits together

Maya / Rhino

Existing design tools

→

Omniverse / OpenUSD

Connective foundation

→

Stable Diffusion + LoRA

Fine-tuned on own data

→

Client-ready renders

Abstract → clear visuals

Where it helps

Handles reformatting, standardization and manual tweaks
Tests thousands of options overnight as an "extra team member"
Keeps designers in their creative flow

Where it falls short

Can't solve construction or program challenges
Shallow grasp of how elements interface
Outputs can be dreamlike yet impractical to build

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.