Home / Case Studies / Life Sciences AI
Case Study · Life Sciences · Private AI

Air-gapped AI for pharmaceutical R&D — zero cloud, 100% on-premises

A 12-person computational chemistry team running drug-candidate screening needed AI compute that never touched a public cloud. Haink supplied NVIDIA DGX Spark systems and RTX 6000 Ada workstations — from quote to powered-on lab in four weeks.

"Our IP counsel was clear: proprietary compound structures cannot leave our perimeter network — not even encrypted to a trusted cloud. We needed hardware, not a cloud contract."

— Head of Computational Chemistry, Pharmaceutical R&D Team

Why on-premises was non-negotiable

🔒IP containment

Drug candidate molecular structures are trade secrets. The team's IP counsel required all compute to remain within the lab's physical perimeter — cloud HSMs were not an acceptable alternative.

Latency for iterative screening

Molecular dynamics simulations run thousands of short jobs per day. Cloud round-trip latency added friction to the iterative loop. Local GPU eliminated the wait entirely.

📊Cost predictability

Cloud GPU spend for this workload varied 3–5× month-to-month as job counts scaled with project phases. Fixed hardware cost was essential for budget forecasting.

🏗️Lab environment compatibility

Desktop-class DGX Spark form factor fits under a lab bench. No raised floor, no precision cooling — operates in standard office conditions up to 30°C.

Hardware supplied

Two NVIDIA product lines — one for personal scientist workstations, one for shared high-throughput compute nodes.

Shared compute nodes

NVIDIA DGX Spark × 4

GPUGB10 Grace Blackwell Superchip
GPU memory128 GB unified (shared CPU+GPU)
CPU72-core NVIDIA Grace (Arm)
System RAM128 GB LPDDR5x
Storage4 TB NVMe SSD
NetworkConnectX-7 100GbE
Form factorDesktop (1.7 kg, no rack required)
Scientist workstations × 6

RTX 6000 Ada workstations

GPUNVIDIA RTX 6000 Ada Generation
GPU memory48 GB GDDR6 ECC
CPUIntel Xeon W-2400 (24-core)
System RAM256 GB DDR5 ECC
Storage2× 4 TB NVMe RAID-0
Network10 GbE to DGX Spark pool
Display4× DP 1.4, molecular viz ready
Interconnect: All 4 DGX Spark units connected at 100GbE via an unmanaged 10/100GbE switch. Each scientist workstation connects to the pool at 10GbE. No internet uplink — isolated VLAN, firewall block at lab edge router.

What changed

4→0.3 wk

Drug candidate screening cycle reduced from ~4 weeks to under 2 days using on-premise DGX compute

4 weeks

Quote to powered-on lab environment, including customs and delivery to the research site

0 bytes

Compound data transferred to cloud infrastructure — full air-gap maintained throughout the project

12 mo

Estimated payback vs equivalent cloud GPU reservation for the same workload profile

Related resources

Case Study H200 Inference Cluster — 9-Day Delivery Cloud break-even in 14 months Case Study 64× H100 SXM5 Sovereign AI Cluster Gulf government entity, licensed export Brand Page NVIDIA hardware at Haink DGX Spark, RTX 6000 Ada, H-series GPUs Private AI Private AI infrastructure overview On-premises AI compute — all scales

Need private AI compute for a research team?

Haink supplies NVIDIA DGX Spark, RTX Pro workstations, and lab-scale inference clusters — no cloud dependency, no data sovereignty compromise.

sales@haink.org