Compute, data, models, integrations. What does a serious Saudi-first AI stack actually look like in 2026 if you build it from the ground up? I’ve been working through this question with MLO Technologies for the better part of a year. Here’s the architecture I’d recommend.

Layer 1 — Compute, hosted in Saudi

Data residency matters in 2026 and matters more in 2027. Three options:

  • HUMAIN — the new Saudi sovereign AI infrastructure operator. Best for serious enterprise workloads.
  • AWS Saudi region — Bedrock + EC2 GPU instances, KSA data residency confirmed.
  • Google Cloud Saudi region — Vertex AI with regional residency option.

Pick based on your existing stack and which models you need. The era of running everything on a generic US-East cluster is ending fast.

Layer 2 — Models, mixed strategy

The right answer is hybrid:

  • Frontier closed (Claude 5, GPT-6) for hardest tasks via region-residency offerings
  • Open-weight Arabic foundation (Allam, Jais, fine-tuned Falcon) for Arabic-native tasks
  • Specialised local fine-tunes for vertical domains (legal, medical, government)

Don’t standardise on one. The frontier-vs-open gap closed enough that hybrid is the default.

Layer 3 — Data and retrieval

  • Vector store with regional residency (Qdrant or Weaviate self-hosted in HUMAIN, or AWS OpenSearch Saudi)
  • Document ingestion pipeline that handles Arabic text correctly (RTL, ligatures, dialect normalisation)
  • Permission boundaries baked in from day one — government and enterprise customers will demand them

Layer 4 — Integration and orchestration

This is where Saudi-first AI gets interesting:

  • MCP servers for every Saudi platform that matters — Salla, Foodics, Mrsool, government APIs
  • Agentic orchestration framework — LangGraph, custom, or Anthropic’s Managed Agents
  • Identity and authentication — Nafath integration for enterprise/government auth flows

MCP is the integration layer for the next decade; build for it now.

Layer 5 — Application surface

The user-facing layer that justifies the rest:

  • Bilingual UX (Arabic-first, English-second)
  • RTL-correct everything
  • Saudi payment rails (Mada, Apple Pay, STC Pay) integrated cleanly
  • Local hosting of all customer-facing assets (Saudi CDN, no US-leakage)

What it costs to start

A minimum viable Saudi-first AI deployment:

  • ~$15K/month compute baseline (scales with usage)
  • ~$50K initial setup (architecture, integrations, fine-tunes)
  • ~$200K/year for a team of 3–4 engineers maintaining it

Real money but not fantasy money. Within reach of a seed-funded company.

The strategic moat

Building Saudi-first creates two compounding advantages:

  1. Government and enterprise lock-in — once you’re inside the data-residency moat, swapping you out is expensive
  2. Arabic-native quality — fine-tunes on local data outperform generic frontier models on Arabic-specific tasks

The principle

The next decade’s biggest AI businesses in MENA won’t be built on a US-default stack. Saudi-first isn’t a constraint; it’s the moat. The earlier you architect for it, the deeper the moat compounds.

Building this stack? Compare architecture notes.