Medical Imaging & Digital Pathology
Highly Differentiated Offering — Where data egress costs exceed compute costs
The Problem in Numbers
Based on internal market research. Figures are directional estimates informed by industry analysis and may vary by organisation.
500–5K
Slides / Day (Mid-Size Pharma)
1–10 GP
Gigapixels per Slide
25 TB
Peak Daily Data Volume
48–72 hrs
Pathologist Review Backlog
Why Cloud Fails for Imaging
Three problems that compound simultaneously:
Latency
A single gigapixel slide takes minutes to upload, wait for inference, and download results. Multiply by thousands of slides per day. Clinical decisions wait.
Compute Cost
Vision model GPU time is expensive. Cloud charges per-second regardless of utilisation. Idle billing compounds with queue wait times.
Data Egress
Every image uploaded and result downloaded incurs bandwidth charges. For imaging, egress costs often exceed compute costs — $30,000 – $80,000 / mo.
Image Types in Scope
Whole-slide images (WSI)
Gigapixel-resolution digital scans of tissue sections used for histopathological diagnosis and biomarker analysis.
Radiology scans
CT, MRI, and X-ray images used for anatomical and functional assessment in clinical trials and diagnostics.
Biomarker images
Immunohistochemistry (IHC) and fluorescence microscopy images used to quantify protein expression and cellular markers.
Companion diagnostic images
Images tied to specific drug therapies — used to determine patient eligibility for targeted treatments.
Manufacturing inspection images
High-resolution images from pharmaceutical production lines for defect detection and quality assurance.
Pathology annotations
Pathologist markups, region-of-interest selections, and diagnostic notes overlaid on digital slides.
The Solution
A four-stage inference platform purpose-built for gigapixel imaging workloads
Image Asset
Management
Management
→
Image
Twin
Twin
→
Diagnostic
Twin
Twin
→
Pathology
Copilot
Copilot
STAGE 1
Image Asset Management
Every image is fingerprinted and catalogued before any AI processing begins. DICOM/SVS format validation, checksum verification, metadata extraction, and version tracking form the file-level control plane for gigapixel assets.
DICOM/SVS Validation
Checksum
Metadata Extraction
Version Tracking
STAGE 2
Image Twin
The raw image is transformed into a machine-readable digital twin. Gigapixel tiling into inference-ready patches, CNN detection/segmentation, VLM annotation, and structured findings — preserving spatial coordinates, confidence scores, and lineage.
Gigapixel Tiling
CNN Detection
VLM Annotation
Structured Findings
STAGE 3
Diagnostic Twin
The image twin is elevated into clinically actionable intelligence — cross-slide correlation, biomarker quantification, pattern detection across patient cohorts, pathology report generation, and flagged anomalies ready for human review.
Cross-Slide Correlation
Biomarker Quant.
Pattern Detection
Report Generation
STAGE 4
Pathology Copilot
Pathologists interact with pre-screened review queues, AI-generated annotations with confidence scores, and guided diagnostic workflows. No diagnostic decision is made without clinical oversight — human-in-the-loop by design.
Review Queues
Confidence Scores
Human-in-the-Loop
Guided Workflows
The CPU bottleneck insight: Tiling a 10-gigapixel image into inference-ready patches is entirely CPU-bound. Without optimisation, the GPU sits idle waiting for patches. Intics solved this exact bottleneck at the Fortune 25 customer — where OCR preprocessing starved GPU inference before optimisation. For imaging, the bottleneck is identical but larger.
Proven Architecture
40–50B
Tokens / Day (Platform Capacity)
124 → 48
Pod Optimisation Journey
NVIDIA Hopper
Vision-Optimised Silicon
The 5-Layer AI Cake
Who delivers what — tailored for imaging workloads
Layer 5: AGENCY (Pathology Copilot)
Review queues, human-in-the-loop, audit trails, diagnostic workflow orchestration, confidence-gated decisions
INTICS
Layer 4: INTELLIGENCE (CNN / VLM Serving)
Silicon-aware placement (NVIDIA Hopper for vision, H100 for multi-modal), batch/stream scheduling, mixed CNN + VLM + OCR pipelines
INTICS
Layer 3: IRON (Servers, Networking & Edge)
GPU-accelerated servers, industrial/edge compute for pathology labs, high-bandwidth switching for image transfer, enterprise storage
HW PARTNER
Layer 2: SILICON (Vision-Optimised)
NVIDIA Hopper (cost-optimised CNN/VLM), H100 (complex multi-modal), Xeon/EPYC (gigapixel tiling & preprocessing)
HW PARTNER
Layer 1: ENERGY (High-Density + Edge)
High-density rack power for sustained CNN workloads, edge power solutions for pathology lab deployment
HW PARTNER
For imaging, the Cake has a unique edge dimension. Pathology labs and manufacturing sites need local inference capability — not round-trips to a central data center. Edge compute extends the captive farm to the point of care.
Financial Impact
H100 Configuration — 48 Pods — with imaging-specific egress savings
$240K
Cloud / Month
$79K
Captive / Month
Mo 8–9
Break-Even
~$5.1–6.9M
3-Year Net Savings
Cost Comparison
| Line Item | Cloud (Hyperscaler) | Owned Infrastructure |
|---|---|---|
| Monthly inference cost | $240,000 / mo | $79,000 / mo |
| One-time hardware Capex | $0 (rented) | $1,750,000 |
| Monthly operations (power, colo, ops team, software, insurance) | Included in rental | $30,000 / mo |
| Data egress (imaging-specific) | $30,000 – $80,000 / mo | $0 |
| GPU idle billing | Billed even when idle | Marginal electricity only |
Financial Outcomes
| Outcome | Value |
|---|---|
| Monthly savings (inference alone) | ~$161,000 / mo |
| Monthly savings (incl. egress elimination) | ~$191,000 – $241,000 / mo |
| Break-even point | Month 8 – 9 |
| 3-year total cloud spend (incl. egress) | $9,720,000 – $11,520,000 |
| 3-year total captive spend | $2,844,000 + $1,750,000 Capex = $4,594,000 |
| 3-year net savings | ~$5,100,000 – $6,900,000 |
| Hardware on balance sheet | $1,750,000 depreciable asset |
The egress kicker: For imaging workloads, data egress costs are often larger than compute costs. Every gigapixel image uploaded and downloaded creates bandwidth charges. Captive inference eliminates this entirely — making the savings case $1–3M stronger than document processing over 3 years.
Time Savings
From days to minutes — across the imaging pipeline
80–85%
Faster Slide Analysis
Days → Min
Pathologist Backlog Cleared
70–80%
Faster Companion Dx
| Process | Before | After | Improvement |
|---|---|---|---|
| Whole-slide image analysis | 15–30 min / slide (cloud round-trip) | 2–5 min / slide (local inference) | 80–85% faster |
| Pathologist review queue | 48–72 hour backlog | Near-real-time AI pre-screening | Days → Minutes |
| Companion diagnostic turnaround | 5–7 days | 1–2 days | 70–80% faster |
| Biomarker quantification (batch) | Days (cost-gated cloud batching) | Hours (continuous local inference) | Days → Hours |
The pathologist backlog reduction is the headline metric. When slides sit in a 48–72 hour queue, clinical decisions wait. Captive inference with AI pre-screening means pathologists review AI-annotated slides in near-real-time — not raw images days later.
Volume & Throughput
Gigapixel-scale inference with vision-optimised silicon
500–5K
Slides / Day
25 TB
Peak Daily Data Volume
NVIDIA Hopper
Vision-Optimised Silicon
40–50B
Platform Token Capacity / Day
| Metric | Capacity |
|---|---|
| Slides per day (mid-size pharma) | 500 – 5,000 |
| Size per whole-slide image | 1 – 10 gigapixels (500MB – 5GB) |
| Daily data volume | 250GB – 25TB |
| Inference models | CNN (detection / segmentation), VLM (report generation), OCR (annotation extraction) |
| Vision-optimised silicon | NVIDIA Hopper — lower cost-per-inference than H100 for CNN/VLM workloads |
| CPU preprocessing (critical bottleneck) | Image tiling, normalisation, quality filtering — CPU-bound |
| Pipeline stages | Image Asset Mgmt → Image Twin → Diagnostic Twin → Pathology Copilot |
| Pod configuration (H100, proven) | 48 pods |
CPU preprocessing is the real bottleneck. Tiling a 10-gigapixel image into inference-ready patches is entirely CPU-bound. Without optimisation, the GPU sits idle waiting for patches. Intics solved this exact bottleneck at the Fortune 25 customer. For imaging, the bottleneck is identical but larger — making CPU optimisation even more critical.
Compliance Framing
AI-assisted diagnostics face the strictest regulatory scrutiny in pharma
FDA 510(k) / De Novo
AI-assisted diagnostic tools require validated, static infrastructure. Cloud auto-scaling invalidates the validation basis. Captive farm runs on fixed, documented hardware that can be submitted as part of the regulatory filing.
Patient Image Data Sovereignty
Diagnostic images contain patient identifiers and clinical information. Must stay within the customer's controlled environment under HIPAA, GDPR, and country-specific regulations. Zero cloud egress.
Reproducibility
Regulatory submissions for AI-assisted diagnostics require identical inference results across runs. Captive infrastructure guarantees hardware consistency. Cloud auto-scaling breaks reproducibility.
Audit Trails
Every image processed, every model prediction, every confidence score, every pathologist review decision — logged immutably for regulatory review.
GxP Namespace Isolation
4-tier environment model (Dev → UAT → Prod → DR) ensures diagnostic production models are completely isolated from development and testing activities.
"FDA inspectors evaluating an AI-assisted diagnostic will ask to see the infrastructure. If it's in the cloud, you cannot demonstrate control. If it's captive, you can walk the inspector through every layer."
Compliance positioning for pharma CIO
Who Delivers What
Your hardware partner provides the bottom 3 layers — with a unique edge angle for imaging. Intics owns the top 2.
Your Hardware Partner — Energy, Silicon, Iron
Energy
Vision inference is power-dense. Sustained CNN workloads generate significant heat. The hardware partner manages thermal design for GPU-heavy racks. Edge deployment at pathology labs requires compact, efficient power solutions.
Silicon
NVIDIA Hopper is the standout silicon for vision workloads — cost-optimised CNN/VLM inference at lower $/image than H100. Xeon/EPYC CPUs are critical for gigapixel tiling, normalisation, and quality filtering.
Iron
GPU-accelerated servers for core inference. Industrial/edge compute for deployment at pathology labs and manufacturing sites. High-speed switching fabric for moving gigapixel images between preprocessing and inference tiers. High-capacity enterprise storage.
Edge deployment is the differentiator. Pathology labs and manufacturing sites need local inference — not round-trips to a central data center. Industrial/edge compute extends the captive farm to the point of care or point of manufacture.
Intics — Intelligence, Agency
Intelligence (Model Serving)
Silicon-aware compute selection — NVIDIA Hopper for cost-optimised vision inference, H100 for complex multi-modal analysis. CPU optimisation for gigapixel tiling is critical and is the real bottleneck. Serving layer handles mixed CNN + VLM + OCR pipelines concurrently.
The CPU bottleneck lesson from the Fortune 25 customer applies even more strongly to imaging — where preprocessing is a larger share of the total pipeline.
Agency (Pathology Copilot)
Multi-model pipeline orchestration across all 4 stages. Job scheduling for batch imaging (overnight processing) vs. urgent cases (stat pathology reviews). Quality gates between pipeline stages. Human-in-the-loop pathologist review integration.
The pathologist gets a pre-screened queue with AI-generated annotations and confidence scores. Intics orchestrates the multi-step pipeline; the pathologist focuses on clinical judgement.
The CPU/GPU optimisation that reduced 100 H100 pods to 48 without performance loss applies even more strongly to imaging — where CPU preprocessing is a larger share of the total pipeline than in document processing.
The Pitch
Direct to the pharma customer
"You are processing thousands of gigapixel pathology slides every day. You are either waiting on cloud round-trips (slow and expensive) or running manual review queues with 72-hour backlogs. There is a third option: own the inference infrastructure, eliminate data egress costs entirely, break even in 9 months, and save up to $6.9M over three years."
Key Numbers to Remember
$5.1–6.9M
3-Year Net Savings
Mo 8–9
Break-Even Point
$0
Data Egress Cost
$240K → $79K
Monthly: Cloud → Captive
80–85%
Faster Slide Analysis
Days → Min
Pathologist Backlog → Pre-Screened
NVIDIA Hopper
Vision-Optimised Silicon
4 Stages
Asset → Image Twin → Dx Twin → Copilot