Cloud vs Self-Hosted AI: 3-Year TCO Calculator (Canada)

Cloud AI APIs offer instant access with no capital outlay. Self-hosted open-weight models offer sovereignty, privacy, and dramatically lower per-token costs at scale. The right answer depends on your workload volume, your hardware budget, your province’s electricity rate, and how much of the clock your inference stack is actually running.

This calculator computes a complete 3-year total cost of ownership for both paths — cloud API billing (token-based or GPU-hour) versus self-hosted hardware — and shows you the exact crossover month. All formulas are shown in the methodology section below. Electricity rates are pre-filled from D-Central’s Canadian electricity rates dataset (June 2026; verify at utility tariff schedule before financial decisions).

We stand on the shoulders of open-source inference projects — llama.cpp, Ollama, vLLM, Unsloth — that made local inference economically viable. None of D-Central’s hardware recommendations require our proprietary software; these frameworks run on any compatible GPU.

3-year AI TCO calculator

Cloud AI pricing reference (as of June 2026)

The prices below are from public pricing pages as of June 2026. API prices change frequently — always verify at the provider’s official pricing page before financial modelling. All prices are in USD per million tokens.

Sources: Anthropic — anthropic.com/pricing (June 2026). OpenAI — openai.com/api/pricing (June 2026). Google — ai.google.dev/pricing (June 2026). Prices change; verify before committing to a budget.

Cloud GPU-hour reference (as of May–June 2026)

Source: Synthesized from BuildMVPFast GPU comparison and SynpixCloud GPU pricing 2026, as of May–June 2026. H100 rental rates have fallen significantly since 2024; verify current rates before committing.

Methodology and formulas

The calculator uses the following formulas, modelled over a 36-month (3-year) horizon. All calculations are performed client-side in your browser; no data is sent to our servers.

Cloud cost

Token-based billing:

monthly_cloud (USD) = (input_MTok × input_price_$/MTok)
                    + (output_MTok × output_price_$/MTok)

monthly_cloud (CAD) = monthly_cloud_USD × USD_CAD_rate

3yr_cloud (CAD) = monthly_cloud_CAD × 36

GPU-hour billing:

monthly_cloud (USD) = gpu_hours_per_month × $/hour

3yr_cloud (CAD)     = monthly_cloud_CAD × 36

Self-host cost

monthly_electricity (CAD) = watts × (utilization% ÷ 100)
                           × 24 h/day × 30.44 days/mo
                           ÷ 1000          {convert Wh → kWh}
                           × $/kWh

monthly_maintenance (CAD) = capex × (annual_rate% ÷ 100) ÷ 12

monthly_opex (CAD)        = monthly_electricity + monthly_maintenance

3yr_self_host (CAD)       = capex + (monthly_opex × 36)

Crossover / payback period

monthly_net_saving = monthly_cloud_CAD − monthly_opex_CAD

IF monthly_net_saving ≤ 0 → cloud is cheaper; no crossover within model
ELSE:
  crossover_months = capex ÷ monthly_net_saving

Assumptions and limitations

• Electricity rate: Rates shown are approximate residential/commercial tariff-schedule values from primary utility sources as of June 2026. Industrial large-power tariffs may differ significantly (lower for >1 MW in Quebec; higher for cryptographic workloads under proposed Hydro-Québec tariffs). See the full electricity rates dataset.
• Currency: Cloud API prices are USD-denominated. Self-host costs (hardware, electricity) are typically CAD. The FX input defaults to 1.37 — update to your transaction-date rate. CAD/USD has ranged from 1.30 to 1.45 in 2025–2026.
• Not modelled: Cloud reserved-instance or committed-use discounts (can reduce cloud cost 20–50 %); colocation rack rent and cooling for self-hosted; enterprise support contracts; network egress charges; insurance; staff hours for hardware administration.
• Maintenance rate: The 10 % default follows the industry rule-of-thumb for server hardware. Consumer GPU workstations may run lower (5–8 %); dense GPU cluster deployments may run higher (12–18 %).
• Hardware lifespan: This model assumes hardware depreciates over 3 years. GPU hardware commonly runs 5+ years; extending the horizon would improve the self-host case further.
• Utilization: The single biggest lever in the model. A GPU running at 20 % utilization is very expensive per useful token. Self-hosting typically makes economic sense above 40–50 % sustained utilization.

Hardware tiers for self-hosted AI in Canada

The local AI hardware guide maps today’s major open-weight models (Gemma 4, Qwen3, Llama 4 Scout, DeepSeek V4 Pro) to verified hardware tiers with VRAM requirements and inference framework recommendations. For enterprise-scale deployments or multi-GPU cluster design, D-Central’s AI sovereignty consulting service covers architecture, procurement, and Hashcenter buildout in Quebec and across Canada.

Key self-hosting hardware brackets to model in the calculator:

• Personal / developer workstation (RTX 4090, 24 GB VRAM): typical system draw ~350–400 W; runs Qwen3-27B at Q4 and smaller models; best for individual developers or small teams with moderate volume.
• Workstation 48 / Apple 128 GB unified: total system draw ~400–600 W; handles Llama 4 Scout INT4 and Qwen3-35B; suitable for teams of 5–20 at sustained workloads.
• Single H100 SXM node: ~700 W GPU draw, ~1.2–1.5 kW total system; production inference server; viable for teams with >100 concurrent users or high-throughput batch pipelines.
• Multi-GPU Hashcenter node (4–8× H100): 4–10 kW total system; enterprise scale; requires colocation or purpose-built facility. Contact D-Central for design.

For capex inputs in the calculator, request a quote through the consulting page — we do not list hardware prices publicly as configurations are built to order.

Frequently asked questions