Internal / appendix-ready

Goodland Cloud Data Service

Technical architecture overview & indicative pricing
Vietnam market entry · Data residency · Storage, backup, IDS, compliance · ERP AI · GPU rental

This document supports the business plan appendix. Diagrams describe a target-state reference architecture. Pricing is illustrative until vendor and tax positions are finalized.

1. Executive alignment

2. High-level architecture (Vietnam boundary)

Clients
SMEs
Enterprises
Tech startups

HTTPS / APIs / Console

Vietnam — processing & data residency
API gateway
& integration APIs
IAM / orgs
RBAC, MFA
Object & file
storage tier
Metadata
& indexing
Backup &
recovery
Encryption
KMS
Managed IDS
SIEM signals
Audit logs
& retention
AI gateway
LLM, RAG, tools
Vector / search
ERP + docs index
GPU pool
train / infer
Monitoring
24/7 NOC
Incident
response
Support
VI / EN

Underlying: in-country data centers or compliant local partner infrastructure (multi-AZ within Vietnam).

3. Layered logical architecture

LayerComponentsRole
IntegrationWeb console, SDKs/CLI, REST & event APIsCustomer access & automation
Control planeIAM, billing/quotas, policy & compliance configurationGovernance & commercial metering
Data planeDistributed object storage, snapshots/versioning, in-VN replicationDurable, scalable data services
Security planeEdge protections (WAF/DDoS as applicable), IDS, hardening baselinesDetective & preventive controls
AI, automation & computeManaged LLM endpoints (e.g. Azure OpenAI, GPT-4o/4.1), RAG pipelines, OCR/Form Recognizer, analytics sandboxes (Python/AutoML), RPA connectors (UiPath, Power Automate), speech (Whisper, Azure Speech), recommendation services, tenant GPU instances (configurable vCPU, RAM, disk, GPU SKU)ERP intelligence, document capture, forecasting, workflows, voice ERP, GPU rental for ML

4. Backup & recovery flow (conceptual)

  1. Customer workload or agent triggers scheduled/immediate backup via API.
  2. Keys obtained from KMS; data encrypted in transit (TLS 1.2+) and at rest.
  3. Chunks + metadata written to storage tier; immutable / retention policies applied by backup service.
  4. IDS and access analytics consume telemetry; anomalies routed to monitoring & IR workflows.
  5. Restore path: verified read, policy checks, decrypt, delivery to customer environment (paths remain in Vietnam).

5. Compliance & security mapping

6. AI applications for internal ERP systems (reference catalog)

Below maps common ERP AI patterns to technology options. Actual stack is selected per tenant (latency, cost, data sovereignty). RAG and knowledge services connect to approved ERP APIs, warehouses, and document stores inside the residency boundary where possible.

#CapabilityPrimary use casesTechnology optionsArchitecture notes
1 Internal AI chatbot (ERP assistant) Natural-language data queries (“this month’s revenue?”), procedure help, instant report Q&A OpenAI GPT-4o / GPT-4.1; Azure OpenAI Service; RAG over ERP DB views, cubes, and curated APIs Orchestration layer enforces RBAC: queries execute as the user’s ERP identity; audit log of questions & tool calls.
2 AI OCR & document processing Invoice → bookkeeping lines; automated data entry from scans/PDFs Google Vision API; Azure AI Document Intelligence / Form Recognizer; custom models on tenant GPUs Pipeline: ingest → classify → extract → validation rules → ERP posting API; human-in-the-loop queue for low confidence.
3 AI data analytics (AI BI) Revenue & inventory forecasting; anomaly detection (e.g. expense spikes); financial commentary Python + ML (scikit-learn); Power BI + AI features; AutoML (Google Vertex / Azure ML) Feature store or warehouse connection; scheduled jobs on CPU or GPU nodes; exports to dashboards and alerts.
4 AI-driven process automation (AI + RPA) Order generation, approval routing, intelligent workflows UiPath (AI Center, Document Understanding); Microsoft Power Automate + AI Builder Event bus from ERP ↔ automation runners; secrets in vault; IDS monitors bot service accounts.
5 AI voice (voice-enabled ERP) Hands-free KPI answers (“today’s revenue?”); mobile & field scenarios Whisper (STT); Azure Speech (STT/TTS); LLM backend as in (1) Low-latency path: edge device → STT → intent → secured data API → TTS response; PII minimization in audio retention.
6 AI personalization (recommendation) Purchase suggestions, pricing hints, lead/customer recommendations Collaborative filtering + gradient boosting; embeddings from product/customer history; optional deep models on GPU Offline training on GPU rental; online serving API with A/B flags and explainability hooks.
7 AI for support Internal L1/L2 bot; ticketing integration; troubleshooting (“why can’t I run AR report?”) Same LLM stack as (1); integration with ITSM/ticket APIs; runbooks as RAG sources Escalation to human with full transcript; link to known-error DB.
8 AI knowledge base Centralized ERP manuals, implementation guides, tribal knowledge Document ingestion, chunking, embeddings, vector DB; optional multilingual models Versioning tied to ERP release; access aligned to module licensing.
9 AI help (contextual process guidance) Step-by-step “create warehouse receipt” with deep links to screens LLM + structured procedure graph; role-aware (accountant vs warehouse); UI context from client shell Differentiator: combines RBAC, current module/screen, and org policies — not static PDFs only.

7. AI / ML platform architecture (conceptual)

Client & ERP tier
Web / mobile ERP
Voice clients
RPA bots

Mutual TLS / OAuth2 / ERP session propagation

Goodland AI services (control + data)
API / agent
orchestration
RAG & tools
ERP SQL/API
OCR & docs
Vision / DI
Analytics
notebooks, AutoML
Speech
STT / TTS
Workflow hub
RPA triggers
Vector DB
manuals, KB
Model registry
versions
Observability
cost, safety

Schedulers, queues, object storage (section 2)

Tenant compute — GPU rental (IaaS)
VM / container
vCPU, RAM
GPU SKU
8–80 GB VRAM
Disk
NVMe / SSD

Customers choose GPU type, VRAM tier, vCPU, RAM, storage, and duration (hourly, monthly reserved). Suitable for training, fine-tuning, batch inference, and self-hosted models.

8. GPU rental — SKU reference (illustrative)

SKUs align to NVIDIA-class accelerators commonly requested for ML. Exact chip generation depends on datacenter supply; VRAM columns reflect customer-facing usable VRAM tiers (8, 16, 24, 32, 40, 48, 80 GB). H100 offered as flagship training/inference tier (typically 80 GB).

TierTypical GPU families (examples)VRAM focusTypical workloads
EntryT4-class, L4-class8–16 GBLight inference, dev/test, small CV/NLP
ProfessionalRTX / A-series workstation class24 GBFine-tuning small LLMs, batch OCR post-processing
High-memoryA30 / L40S-class (examples)32–48 GBLarger models, multi-GPU optional
DatacenterA100-class40 / 80 GBTraining, LLM serving at scale
FlagshipNVIDIA H10080 GBFrontier training, large-scale inference, HPC-style jobs

Add vCPU and system RAM in standard blocks (e.g. 8–32 vCPU, 32–256 GB RAM) and network-attached or local NVMe volumes per performance tier.

9. Indicative pricing — core data platform (VND / month)

Price type: These numbers are indicative B2B list / retail prices to end customers (what you charge businesses), not Goodland’s wholesale cost from AWS or partners, and not distributor buy prices unless you publish a separate channel list. For COGS, hyperscaler wholesale, allocated 24/7 support, Vietnam overhead, and competitor list benchmarks, use Goodland-Cloud-Unit-Economics-Benchmarking (HTML/PDF).

VAT may apply. Annual prepay typically 15% below 12× monthly on Starter–Business tiers. Enterprise is custom.

Tier Target segment Included storage Backup IDS / monitoring Support Indicative monthly (VND)
Starter Small SME, pilots 500 GB Daily, 14-day retention Basic alerting Business hours, ticket 2,500,000 – 4,000,000
Growth Growing SME, startups 2 TB Daily + weekly, 30-day Managed IDS bundle, monthly report Extended + chat 8,000,000 – 12,000,000
Business Mid-market 10 TB Policy-based RPO/RTO options Dedicated correlation, SOC integration option 24/7, optional CSM 28,000,000 – 45,000,000
Enterprise Regulated / large orgs Custom (50 TB+) In-VN geo-redundancy, legal hold Custom playbooks, IR retainer 24/7 + on-site (major cities) Custom (from ~80,000,000 + commit)

Add-ons — core platform (illustrative)

Add-onNotesTypical range (VND / month)
Extra storagePer TB, volume discounts400,000 – 900,000 / TB
Extended retentione.g. 90d → 365d+15–35% on backup component
API & integration packHigher limits, dedicated endpoints1,500,000 – 5,000,000
Premium SLAe.g. 99.9% target+20–40% platform fee

10. Indicative pricing — ERP AI services (VND)

Bundles assume Goodland-hosted orchestration + metering. Third-party model usage (OpenAI, Azure OpenAI, Google Cloud) is often passed through at cost + margin or requires customer bring-your-own-key (BYOK). Numbers are order-of-magnitude for planning.

PackageWhat’s includedIndicative monthly (VND)Notes
AI ERP LiteInternal assistant + knowledge base (1), (8); up to ~5k RAG queries/mo; 1 connector6,000,000 – 12,000,000Add seats or queries ala carte
AI ERP StandardLite + OCR pipeline (2) up to 2k pages/mo + support bot (7)15,000,000 – 28,000,000Extra pages 800 – 2,500 VND/page by volume
AI ERP PlusStandard + voice channel (5) + workflow hooks for RPA (4) — 2 flows28,000,000 – 48,000,000STT/TTS minutes billed separately (below)
AI Analytics add-onForecasting & anomaly jobs (3); 1 dashboard workspace; scheduled scoring8,000,000 – 22,000,000Excludes Power BI / cloud AutoML licenses if customer-owned
Recommendations add-onReco API (6); retrain monthly; A/B hooks5,000,000 – 15,000,000Heavy training uses GPU hours (section 11)
Contextual AI HelpRole-aware guidance (9); procedure graph build + 40h professional services once4,000,000 – 10,000,000 / mo + one-time 40,000,000 – 120,000,000PS for screen-map & content curation

AI usage meters (illustrative)

MeterUnitTypical range (VND)
LLM / RAG query (after bundle)per 1,000 calls600,000 – 2,500,000
STT (Whisper / Azure Speech)per audio hour180,000 – 550,000
TTSper 1M characters400,000 – 1,200,000
OCR / document AIper page800 – 2,500
RPA production botper bot / month3,000,000 – 9,000,000

11. Indicative pricing — GPU rental (VND / hour)

Linux VM with selected GPU. vCPU and RAM priced additively (examples: 2,000 – 8,000 VND / vCPU-hour; 1,500 – 5,000 VND / GB-RAM-hour). Disk: ~400 – 1,200 VND / GB-month (NVMe premium). Committed use: 1-month commit ~8–12% off; 12-month ~20–35% off headline hourly. Spot / interruptible (if offered): ~40–70% below on-demand. VAT may apply.

GPU tierVRAM (customer-facing)Indicative VND / GPU-hourComment
Entry (e.g. T4 / L4 class)8 GB12,000 – 28,000Dev / light inference
Entry+16 GB18,000 – 38,000Small fine-tunes
Workstation-class24 GB35,000 – 75,000Mid-size models
High-memory32 GB55,000 – 110,000Training / batch
High-memory40–48 GB85,000 – 190,000Larger batches, multi-worker
Datacenter (A100-class)40 GB150,000 – 320,000Production training
Datacenter (A100-class)80 GB220,000 – 450,000Large-model training
Flagship (H100-class)80 GB480,000 – 980,000Frontier training & heavy inference

Packaging: Publish bundled SKUs (GPU + fixed vCPU/RAM/disk) for common sizes to simplify quoting; use component meters above for custom builds.

Goodland Cloud — Architecture & Pricing appendix (core platform, ERP AI catalog, GPU rental). Generated for planning; not a contractual offer. GPU and model prices fluctuate with supply and FX. © Goodland Cloud (concept document).
Related: implementation & support — Goodland-Cloud-Technical-Overview-Support; COGS, benchmarks & margins — Goodland-Cloud-Unit-Economics-Benchmarking (HTML/PDF).