DeepSeek vs ChatGPT: What Cheap Open-Weight AI Actually Changes for a Canadian Business
The DeepSeek and ChatGPT comparison is usually written as a benchmark race, and benchmark leadership changes several times a year. For a business owner deciding what to actually deploy, the ranking on a leaderboard is close to irrelevant by the time a decision gets made.
The difference that persists is structural. ChatGPT is a hosted service: you send data to a provider and receive an answer back. DeepSeek’s models were released with open weights, meaning the model file itself can be downloaded and run on hardware you control. Every consequence that matters commercially, including where your client data goes, what you can promise in a contract, and how the spending is treated for tax, follows from that one distinction rather than from capability.
The comparison that survives the news cycle
| Hosted service | Self-hosted open-weight model | |
|---|---|---|
| Where data goes | To the provider’s infrastructure | Stays wherever you run it |
| Cost shape | Per-token or per-seat, scales with use | Hardware and operating cost, largely fixed |
| Setup effort | Minutes | Real engineering time |
| Capability at the frontier | Usually ahead | Usually behind, and closing |
| Updates | Automatic, and sometimes unannounced | You choose when, and can pin a version |
| Reproducibility | The model can change under you | A pinned model gives the same answer next year |
| Vendor failure risk | Service disappears, capability with it | The weights you downloaded still run |
| Tax treatment of spending | Usually a current expense | Hardware is capital, software may be too |
| Suitable for identifiable client data | Only with the right contractual terms | Yes, if the hardware is yours |
The row that decides most professional-services cases is reproducibility. If you have to explain in two years why a figure was calculated a particular way, a hosted model that has been silently updated four times since cannot reproduce its own output. A pinned local model can.
Data residency is not an abstraction in Canada
Two obligations bite here, and neither is satisfied by a vendor’s security marketing.
Where the records have to live
Records must generally be kept in Canada. The CRA’s position in IC05-1R1 is that records held on servers outside Canada and merely accessed from here are not records kept in Canada. Permission to keep electronic records elsewhere can be requested from your tax services office, on conditions about availability and format. This is about books and records rather than about a chat transcript, but a workflow where the AI tool becomes the system of record crosses the line quietly.
Sending client data is a disclosure
Sending client information to a processor is a disclosure. It is governed by PIPEDA like any other transfer to a third party. Whether the provider trains on your inputs, where processing happens, and what the retention terms say are questions you have to be able to answer to a client who asks. If the terms are silent on training, treat that as a no.
Why regulated practices run models locally
For an accounting practice, a law firm, a medical clinic or anyone holding regulated client data, this is the entire argument for running a model locally. It is not about cost. It is that no data leaves the building, so there is no disclosure to consent to and no cross-border question to answer. The same argument runs on consumer hardware, where processing on the device rather than in a data centre removes the residency question instead of documenting it. The related security practice is in cyber security for small business finance.
How the spending is treated
The tax treatment splits along the same line as the deployment choice, which is worth modelling before committing.
Subscriptions and API usage are ordinary operating costs, deductible in the year under the tests in what the CRA allows as a business expense: incurred to earn income, and reasonable in the circumstances. Simple, and usually the right answer for a small business.
Hardware is capital. General-purpose electronic data processing equipment and its systems software falls in Class 50 at a 55% declining balance rate. Computer software that is not systems software falls in Class 12 at 100%, subject to the half-year rule for that category. An enhanced first-year deduction has been proposed for Class 50 additions acquired after 15 April 2024 that become available for use before 2027, and the status of incentive measures should be confirmed rather than assumed. The mechanics are in capital cost allowance explained.
The practical effect is that buying a GPU server to run an open-weight model converts a fully deductible monthly cost into a capital asset written off over several years. That is not a reason to avoid it. It is a reason to run the comparison on after-tax cash flow rather than on the sticker prices.
The SR&ED question, answered honestly
The most common misconception in Canadian AI spending is that adopting AI is research and development. It usually is not.
To qualify for SR&ED, work must be conducted in Canada, must be undertaken for the advancement of scientific knowledge or to achieve a technological advancement, and must be a systematic investigation carried out by experiment or analysis. Routine work is excluded, including quality control and routine data collection carried out to support normal business operations.
Applying that test:
- Subscribing to an AI service and using it: not SR&ED
- Fine-tuning an open-weight model on your own documents using documented, established techniques: usually routine adaptation, not SR&ED
- Encountering a genuine technological obstacle that published knowledge does not resolve, and running systematic experiments to overcome it: potentially SR&ED, if it is documented as it happens
The determining factor is technological uncertainty and a systematic investigation of it, evidenced contemporaneously. A claim assembled after the fact from a commit history rarely survives review. The program is covered in SR&ED tax credits for Canadian businesses.
Choosing, for a small Canadian business
Most businesses should start hosted and stay hosted. The setup cost of running models yourself is real, the capability at the frontier is generally hosted, and a subscription is a deductible operating cost with no capital commitment.
The case for self-hosting turns on three conditions, and it is strong when two or more apply:
- Client data cannot leave your control for regulatory, professional or contractual reasons.
- Volume is high and predictable, so a fixed cost beats a per-use one.
- Reproducibility matters, because you have to defend an output later.
The case is weak when usage is occasional, when you need the current frontier, or when nobody in the business wants to operate infrastructure. Paying a subscription so that somebody else runs the servers is a perfectly good answer.
What the competition actually delivered
The durable result of a credible open-weight competitor is not that one product won. It is that the price of adequate capability fell, and that running a capable model on your own hardware stopped being exotic. For a Canadian professional practice that is the more consequential development, because it turns the data residency problem from a compliance obstacle into an implementation detail.
The corresponding caution is that the tool has never carried the liability. Whatever produced a number, a person signs the return and a person is accountable for the position, which is set out in AI in accounting and tax. No comparison of models changes that, and no vendor has ever received a notice of reassessment.
If you are weighing an AI subscription against buying hardware, and want the after-tax cost and the record-keeping obligations modelled before you commit, that is a straightforward conversation. The local picture is in AI in accounting in Ottawa.
More on technology
Written by