Local AI or public cloud: the complete business case to justify your choice

“Local AI or public cloud” covers two different decisions, and mixing them derails the discussion. The first is a tool decision: whether or not to give your employees ChatGPT, Copilot or Claude subscriptions. Our subscription comparison covers it in detail, verified prices included. The second is an infrastructure decision: where the artificial intelligence that will read your quotes, your contracts and your production data will run. That one commits the company for years, and it is the one this article equips you for.

By local AI, we mean an open-weight model (Qwen, Gemma, Mistral and company, see our model comparison) running on a server you own, inside your walls or at a Canadian-owned host in Quebec. By public cloud, we mean the APIs and subscriptions of the large American providers, whatever the geographic region of their servers: we will see why that nuance matters more than everything else.

What follows is a business case, not a closing argument: every figure is attributed to its source and the last section lists the cases where the public cloud remains the right choice. A case that hides its weaknesses does not survive the first question from a board of directors.

Law 25 requires a privacy impact assessment before personal information is communicated outside Quebec (section 17 of the private-sector act, CQLR c. P-39.1) and before any information system project that processes it (section 3.3). Sending your customers’ emails, your employee files or your quotes into an American provider’s API is precisely such a communication: the assessment must demonstrate that the information would receive adequate protection. It can be done, but it is a compliance file to assemble, defend and keep current. A model running on your server in Quebec makes the transfer question disappear: our PIA guide shows the difference on the documents themselves.

The 2018 U.S. CLOUD Act adds the decisive argument: it compels providers subject to American law to hand over the data in their custody to the authorities, no matter where it is hosted. An American cloud giant’s Montreal region therefore remains under American jurisdiction, because what counts is the owner, not the data centre’s postal code. Our CLOUD Act analysis details the mechanics. A server you own answers to no foreign provider: the immunity comes from the architecture, not from a contract clause.

The penalties give the argument its weight: Law 25 provides for administrative monetary penalties of up to $10 million or 2% of worldwide revenue and penal fines of up to $25 million or 4%. Before a board, the question “why invest in a server?” then turns around: why accept a recurring legal risk when the alternative is an asset?

Let us lay out the figures, all verified on the vendors’ official pages on July 16, 2026 in our subscription comparison. ChatGPT Business costs CA$27 per user per month billed annually; Microsoft 365 Copilot costs CA$40.70, on top of an eligible Microsoft 365 licence; Claude Team costs US$20 per standard seat. Over three years and 30 users, without a single price increase, that comes to around CA$29,000 for ChatGPT Business and around CA$44,000 for Copilot before the base licence. At 100 users, multiply by more than three. And at the end of those three years, nothing remains: no hardware, no software, no accumulated knowledge that belongs to you.

On the other side, a server capable of running a 27B-class model, our default recommendation for an SME, can be found under $10,000 according to our model comparison. That hardware can qualify for Revenu Québec’s C3I tax credit (15% to 25% depending on the region, on the portion above the $5,000 threshold). Putting a first agent into production (the software, the integration with your tools, the commissioning) comes on top; that is what turns the server into a business result, and it is a fundable project: our overview of Quebec funding details ESSOR, NRC IRAP and the other levers that regularly cover 30% to 50% of an eligible project.

The accounting difference matters as much as the amounts: the subscription is a perpetual operating expense, indexed to the provider’s price increases and to your headcount; the server and its agents are a depreciable capital asset, whose marginal cost per additional user is close to zero. A growing business watches the first bill swell every year and the second spread ever thinner.

The argument least often made to executive teams is nonetheless the most concrete. A generic subscription knows nothing of your procedures, your quote templates or your project history: every employee starts from scratch in a chat window. A local AI learns your documents through RAG, plugs into your systems (email, accounting, ERP, document management) and executes your processes instead of commenting on them. At Les Industries Romy, a structural steel fabricator in Granby, shop drawings, quotes and accounts payable now run on the company’s own server, under human validation.

Governance follows the same logic. On your server, every query and every answer can be logged on your premises, access is controlled by your own rules and personal information is redacted before any processing. The terms of service do not shift under your feet: nobody can unilaterally amend the contract of a server you own. And reversibility is observed rather than negotiated: models under the Apache 2.0 licence, documented configurations, data that never left your walls.

That leaves dependency: a business whose critical processes rest on a third party’s API adds a link it controls neither technically nor commercially. The provider can withdraw a model, change its prices or degrade a version without meaningful notice. It is an ordinary business continuity risk, and it is managed like the others: by taking it out of the equation when the stakes justify it.

First, raw quality. The largest cloud models remain superior to SME-sized open models on the hardest tasks: long reasoning, complex code, open-ended questions outside your documents. For a business document assistant, context closes most of the gap: an open model that reads your documents beats a giant that has never seen them. But the claim “local equals cloud at everything” would not survive a demonstration, so do not make it.

Second, the initial investment and the skills. The cloud starts up within an hour with no capital outlay; local requires a server, an integration and someone to maintain it, in-house or through a partner. Managed maintenance exists precisely for that, but it has to be budgeted.

Third, measuring the gain. In April 2026, Statistics Canada published an analysis that puts expectations in their place: the raw productivity gap between AI-using firms and the others (16.8%) shrinks to 5.1% once complementary investments in cloud computing, R&D and training are taken into account. In other words, neither local nor cloud produces a gain by mere presence: the complete project, clean data and training included, is what produces it. Our ROI measurement method draws the practical consequences.

Finally, hybrid is often the real answer: sensitive data and business processes on the local server, generic tasks with no personal information (translating a public document, a general search) on a subscription. The business case does not aim to ban the cloud; it aims to keep your company data from passing through it by default.

Here is the business case condensed, in the order in which it convinces. Every line points back to a verifiable fact in this article or its sources.