// Private AI South Africa — Local AI pricing and recommendation data. // Keep volatile prices here so the public guide can be refreshed without // rewriting page components. const LOCAL_AI_LAST_VERIFIED = '17 July 2026'; const LOCAL_AI_TIERS = [ { id: 'private-office', num: '01', label: 'Private office', name: 'Private Office Node', infrastructureCost: 'R60k–R85k', hardwareCost: 'R46k–R53k base hardware', hardware: 'Mac mini M4 Pro · 64GB unified memory', capacity: '64GB', power: 'Low draw', users: '1–5 users', modelClass: '35B-class Q4 text models', summary: 'A quiet, low-power node for private document search, drafting and internal knowledge work in a small office or branch.', bestFor: ['Legal and accounting', 'Consulting teams', 'Private document assistants'], }, { id: 'team-appliance', num: '02', label: 'Team appliance', name: 'Team AI Appliance', infrastructureCost: 'R130k–R175k', hardwareCost: 'R104,999 base hardware', hardware: 'NVIDIA DGX Spark · 128GB unified memory · 4TB', capacity: '128GB', power: 'Efficient', users: '3–15 users', modelClass: 'Large local models and NVIDIA AI tooling', summary: 'A compact Linux and NVIDIA appliance for shared private AI, larger models, prototyping, data science and edge-development work.', bestFor: ['Clinics and universities', 'Operations knowledge', 'AI and edge development'], }, { id: 'throughput', num: '03', label: 'Throughput', name: 'Throughput Workstation', infrastructureCost: 'R150k–R220k', hardwareCost: 'R64k–R75k GPU budget', hardware: 'NVIDIA GeForce RTX 5090 · 32GB VRAM', capacity: '32GB VRAM', power: 'High draw', users: '5–30 users', modelClass: 'Fast small-to-medium models, voice and image', summary: 'The speed-first option for classification, customer support, transcription, image generation and other high-volume workloads.', bestFor: ['Contact centres', 'Customer operations', 'Media and engineering'], }, { id: 'enterprise', num: '04', label: 'Enterprise', name: 'Enterprise AI Node', infrastructureCost: 'R360k–R550k', hardwareCost: 'R243,999 GPU budget', hardware: 'NVIDIA RTX PRO 6000 Blackwell · 96GB ECC', capacity: '96GB ECC', power: 'High draw', users: 'Shared service', modelClass: 'Large models, controlled concurrency and ECC', summary: 'A governed workstation or server node when larger VRAM, ECC memory, warranty coverage and controlled multi-user inference justify the premium.', bestFor: ['Banks and insurers', 'Government and SOEs', 'Regulated enterprise'], }, ]; const PRIVATE_AI_HIGHLIGHTS = [ { num: '01', label: 'Controlled boundary', value: 'Approved models, documents and integrations run on infrastructure your organisation controls.', }, { num: '02', label: 'South African budgets', value: 'Four infrastructure starting points span R60k to R550k before AT professional services.', }, { num: '03', label: 'Workload-fit compute', value: 'Choose memory for larger text models or GPU throughput for voice, image and repeat processing.', }, { num: '04', label: 'Operational by design', value: 'Power resilience, identity, audit, backups and human review are planned with the machine.', }, ]; const PRIVATE_AI_RUN_PROFILES = [ { id: 'documents', label: 'Private documents', prompt: 'Search, drafting and cited answers across a bounded private corpus.', tierId: 'private-office', modelClass: '20B–35B quantised text model', rationale: 'Prioritise usable memory, quiet operation and straightforward administration over maximum token throughput.', operatingCheck: 'Matter-level permissions, source citations and human approval.', }, { id: 'shared-knowledge', label: 'Shared team knowledge', prompt: 'A multi-user assistant, research environment or internal knowledge service.', tierId: 'team-appliance', modelClass: '35B–70B quantised or mixture-of-experts model', rationale: 'The larger shared-memory pool gives a team room to test more capable models and NVIDIA tooling on one appliance.', operatingCheck: 'Identity, concurrency tests, retrieval evaluation and a named service owner.', }, { id: 'throughput-work', label: 'Voice, image + volume', prompt: 'Transcription, classification, image generation or repeat customer-support tasks.', tierId: 'throughput', modelClass: '9B–27B fast model plus specialised media pipelines', rationale: 'A high-throughput GPU is more valuable here than fitting the largest possible text model into memory.', operatingCheck: 'Queue limits, escalation paths, quality thresholds and power capacity.', }, { id: 'governed-service', label: 'Governed enterprise service', prompt: 'Controlled multi-user inference for regulated or operationally critical workflows.', tierId: 'enterprise', modelClass: '70B–120B quantised model with controlled concurrency', rationale: 'ECC capacity and an enterprise host support larger models, stronger isolation and an operational service boundary.', operatingCheck: 'SSO, audit logging, network isolation, backups and formal accountability.', }, ]; // Condensed Pareto-frontier benchmark points sourced from local.ai on the // verification date below. These are measured source values, not AT estimates. // South African infrastructure budgets remain in LOCAL_AI_TIERS above. const PRIVATE_AI_BENCHMARKS = { attributionLabel: 'local.ai', attributionUrl: 'https://local.ai/#intelligence-speed-frontier', retrieved: '17 July 2026', intelligenceFormula: '100 × (0.10 × τ²-bench + 0.35 × GAIA + 0.55 × GDPval)', taskSpeedFormula: '3,600 ÷ measured task time', valueFormula: '10,000,000 × 2^(intelligence ÷ 10) ÷ (task time × ZAR hardware price)', priceFallbackFormula: 'USD source price × 1.15 VAT × 1.25 AT margin × R18/USD', tiers: { 'private-office': { hardwareId: 'm4_pro_64_20c', hardwareName: 'Mac mini M4 Pro 64GB · 20-core GPU', shortName: 'M4 Pro 64GB', imagePath: 'assets/private-ai/m4-pro-mac-mini.png', imageAlt: 'Mac mini M4 Pro hardware', imageScale: 0.82, priceUsd: 2250, localPriceZar: 49500, localPriceLabel: 'R46k–R53k', memoryGb: 64, measuredSweeps: 43, modelIdentities: 43, frontier: [ { model: 'gemma-4-e2b-q4kxl', quant: 'q4kxl', intelligence: 17.6, taskSeconds: 45.7, tasksPerHour: 78.8 }, { model: 'gemma-4-e2b-q8kxl', quant: 'q8kxl', intelligence: 17.9, taskSeconds: 59.6, tasksPerHour: 60.4 }, { model: 'gemma-4-e4b-q4km', quant: 'q4km', intelligence: 22.8, taskSeconds: 72.6, tasksPerHour: 49.6 }, { model: 'holo-3.1-35b-a3b-q4km', quant: 'q4km', intelligence: 59.1, taskSeconds: 211.5, tasksPerHour: 17.0 }, { model: 'ornith-1.0-35b-q4km', quant: 'q4km', intelligence: 72.7, taskSeconds: 241.1, tasksPerHour: 14.9 }, { model: 'qwen3.6-27b-mtp-q4km', quant: 'q4km', intelligence: 73.1, taskSeconds: 1542.0, tasksPerHour: 2.3 }, ], }, 'team-appliance': { hardwareId: 'gb10_121', hardwareName: 'NVIDIA DGX Spark 128GB', shortName: 'DGX Spark', imagePath: 'assets/private-ai/dgx-spark-single.png', imageAlt: 'NVIDIA DGX Spark hardware', imageScale: 1.08, priceUsd: 4699, localPriceZar: 104999, localPriceLabel: 'R104,999', memoryGb: 128, measuredSweeps: 291, modelIdentities: 103, frontier: [ { model: 'gemma-4-e2b-q4kxl', quant: 'q4kxl', intelligence: 17.6, taskSeconds: 31.4, tasksPerHour: 114.5 }, { model: 'gemma-4-e4b-q4km', quant: 'q4km', intelligence: 22.8, taskSeconds: 57.0, tasksPerHour: 63.2 }, { model: 'gemma-4-E4B-it-FP8-Dynamic', quant: 'fp8', intelligence: 24.4, taskSeconds: 91.8, tasksPerHour: 39.2 }, { model: 'Holo-3.1-35B-A3B-NVFP4', quant: 'nvfp4', intelligence: 61.0, taskSeconds: 131.9, tasksPerHour: 27.3 }, { model: 'Qwen3.6-35B-A3B-NVFP4', quant: 'nvfp4', intelligence: 74.1, taskSeconds: 232.9, tasksPerHour: 15.5 }, { model: 'Qwen3.6-35B-A3B-FP8', quant: 'fp8', intelligence: 74.4, taskSeconds: 298.0, tasksPerHour: 12.1 }, { model: 'Qwen3.6-35B-A3B', quant: '', intelligence: 75.5, taskSeconds: 514.6, tasksPerHour: 7.0 }, { model: 'step-3.7-flash-iq4xs', quant: 'iq4xs', intelligence: 75.6, taskSeconds: 605.9, tasksPerHour: 5.9 }, { model: 'Step-3.7-Flash-148B', quant: '', intelligence: 78.8, taskSeconds: 896.4, tasksPerHour: 4.0 }, ], }, throughput: { hardwareId: 'rtx5090_32', hardwareName: 'NVIDIA GeForce RTX 5090 32GB', shortName: 'RTX 5090', imagePath: 'assets/private-ai/rtx-5090.png', imageAlt: 'NVIDIA GeForce RTX 5090 graphics card', imageScale: 1.04, priceUsd: 3600, localPriceZar: 69500, localPriceLabel: 'R64k–R75k', memoryGb: 32, measuredSweeps: 54, modelIdentities: 27, frontier: [ { model: 'LFM2.5-8B-A1B', quant: '', intelligence: 14.6, taskSeconds: 7.1, tasksPerHour: 509.7 }, { model: 'gemma-4-e2b-q4kxl', quant: 'q4kxl', intelligence: 17.6, taskSeconds: 9.3, tasksPerHour: 388.2 }, { model: 'gemma-4-e2b-q8kxl', quant: 'q8kxl', intelligence: 17.9, taskSeconds: 12.0, tasksPerHour: 299.6 }, { model: 'gemma-4-e4b-q4km', quant: 'q4km', intelligence: 22.8, taskSeconds: 13.8, tasksPerHour: 260.7 }, { model: 'gemma-4-E4B-it-FP8-Dynamic', quant: 'fp8', intelligence: 24.4, taskSeconds: 17.9, tasksPerHour: 201.3 }, { model: 'gpt-oss-20b', quant: '', intelligence: 43.4, taskSeconds: 32.5, tasksPerHour: 110.8 }, { model: 'Gemma-4-26B-A4B-NVFP4', quant: 'nvfp4', intelligence: 63.0, taskSeconds: 67.5, tasksPerHour: 53.3 }, { model: 'gemma-4-31b-q4km', quant: 'q4km', intelligence: 65.6, taskSeconds: 82.3, tasksPerHour: 43.7 }, ], }, enterprise: { hardwareId: 'rtxpro6000_96', hardwareName: 'NVIDIA RTX PRO 6000 Blackwell 96GB', shortName: 'RTX PRO 6000', imagePath: 'assets/private-ai/rtx-pro-6000.png', imageAlt: 'NVIDIA RTX PRO 6000 Blackwell graphics card', imageScale: 1.04, priceUsd: 13250, localPriceZar: 243999, localPriceLabel: 'R243,999', memoryGb: 96, measuredSweeps: 122, modelIdentities: 66, frontier: [ { model: 'Laguna-XS-2.1-NVFP4', quant: 'nvfp4', intelligence: 1.7, taskSeconds: 6.7, tasksPerHour: 540.5 }, { model: 'gemma-4-e2b-q4kxl', quant: 'q4kxl', intelligence: 17.6, taskSeconds: 10.4, tasksPerHour: 346.3 }, { model: 'Qwen3.5-9B-AWQ-4bit', quant: 'awq-4bit', intelligence: 43.2, taskSeconds: 28.9, tasksPerHour: 124.7 }, { model: 'gpt-oss-20b', quant: '', intelligence: 43.4, taskSeconds: 32.1, tasksPerHour: 112.2 }, { model: 'holo-3.1-35b-a3b-q4km', quant: 'q4km', intelligence: 59.1, taskSeconds: 36.3, tasksPerHour: 99.2 }, { model: 'ornith-1.0-35b-q4km', quant: 'q4km', intelligence: 72.7, taskSeconds: 44.3, tasksPerHour: 81.3 }, { model: 'Qwen3.6-35B-A3B-NVFP4', quant: 'nvfp4', intelligence: 74.1, taskSeconds: 53.0, tasksPerHour: 67.9 }, { model: 'Qwen3.6-35B-A3B-FP8', quant: 'fp8', intelligence: 74.4, taskSeconds: 64.1, tasksPerHour: 56.2 }, { model: 'Qwen3.6-35B-A3B', quant: '', intelligence: 75.5, taskSeconds: 91.3, tasksPerHour: 39.4 }, { model: 'Qwen3.6-27B-NVFP4', quant: 'nvfp4', intelligence: 76.4, taskSeconds: 156.9, tasksPerHour: 22.9 }, ], }, }, }; const LOCAL_AI_INDUSTRIES = [ { id: 'legal', label: 'Legal + finance', icon: 'shield', tierId: 'private-office', infrastructureBudget: 'R60k–R85k', workload: 'Contracts, policies, private document search, drafting and cited internal answers.', reason: 'Unified memory accommodates capable quantised text models without the noise, power and administration of a GPU tower.', safeguard: 'Human approval, matter-level permissions, source citations and auditable document versions.', }, { id: 'health', label: 'Healthcare', icon: 'signal', tierId: 'team-appliance', infrastructureBudget: 'R140k–R180k', workload: 'Local transcription, administrative assistance and retrieval across approved clinical documentation.', reason: 'A shared appliance keeps sensitive workloads close to the practice while supporting larger local models and voice tooling.', safeguard: 'No autonomous diagnosis; encrypt records, restrict roles and keep a qualified person responsible for every clinical decision.', }, { id: 'retail', label: 'Retail + support', icon: 'users', tierId: 'throughput', infrastructureBudget: 'R170k–R250k', workload: 'Customer-service retrieval, product matching, call transcription and repeat classification at volume.', reason: 'The workload benefits more from fast small-to-medium models and accelerated voice processing than from maximum model size.', safeguard: 'Confidence thresholds, escalation paths, transcript retention rules and continuous answer-quality evaluation.', }, { id: 'industry', label: 'Industry + field', icon: 'cube', tierId: 'team-appliance', infrastructureBudget: 'R150k–R300k', workload: 'Maintenance knowledge, field guidance, computer-vision development and offline operational assistance.', reason: 'NVIDIA tooling and 128GB memory make the appliance a useful development hub before workloads move to dedicated edge devices.', safeguard: 'Keep safety-critical control outside the language model and validate every edge workflow under real field conditions.', }, { id: 'creative', label: 'Media + engineering', icon: 'spatial', tierId: 'throughput', infrastructureBudget: 'R150k–R220k', workload: 'Creative agents, transcription, image generation, media processing and engineering knowledge assistance.', reason: 'A CUDA workstation offers the broadest compatibility for image, video, voice and accelerated production tooling.', safeguard: 'Track source rights, generated-asset provenance, client confidentiality and model licensing.', }, { id: 'education', label: 'Education + research', icon: 'grow', tierId: 'team-appliance', infrastructureBudget: 'R140k–R450k', workload: 'Shared AI teaching, coding assistance, research experiments and private institutional knowledge.', reason: 'One shared appliance can serve a pilot lab; additional private-office nodes can expand hands-on access later.', safeguard: 'Test student-facing accuracy, disclose AI use and preserve academic assessment integrity.', }, { id: 'regulated', label: 'Regulated enterprise', icon: 'terminal', tierId: 'enterprise', infrastructureBudget: 'R360k–R550k', workload: 'Controlled multi-user inference, larger private models and audited knowledge workflows.', reason: 'ECC VRAM, enterprise-class capacity and a governed host are appropriate when the service becomes operational infrastructure.', safeguard: 'POPIA-aligned purpose controls, SSO, audit logging, network isolation, tested backups and a formal operating owner.', }, ]; const localAiTierById = (id) => LOCAL_AI_TIERS.find((tier) => tier.id === id); Object.assign(window, { LOCAL_AI_LAST_VERIFIED, LOCAL_AI_TIERS, PRIVATE_AI_HIGHLIGHTS, PRIVATE_AI_RUN_PROFILES, PRIVATE_AI_BENCHMARKS, LOCAL_AI_INDUSTRIES, localAiTierById, });