Open-weight versus proprietary models, a workload based Canadian cost model

What follows is a model, not a price list. Prices were observed 12 September 2026 and will have moved. The structure moves more slowly, and the structure is what decides.

Hosted proprietary frontier models price output tokens at roughly five to six times input tokens. The ratio is stable across providers and across tiers within a provider, which suggests a deliberate pricing architecture rather than an accident. Open-weight models served through managed interfaces price output at roughly one to four times input. At least one prices input and output identically.

That difference is the whole article, because it means the cost gap between the two classes is not constant across workloads. It widens sharply for generation heavy work, drafting long documents, writing code, producing summaries of summaries. It narrows for input heavy work, which is what document question answering is: a large amount of retrieved context in, a short grounded answer out.

So the first question is the shape of your traffic, before the class of model. Retrieval over your own corpus sits where the two classes are closest. Long form generation at volume sits where they are furthest apart.

Take a representative document question answering unit of 10,000 input tokens and 1,500 output tokens. That shape is a choice, not a standard, and it is the one to argue with first if you disagree with the conclusion.

In United States dollars, that one request costs roughly 8.75 cents at mid tier proprietary frontier pricing: five dollars per million input tokens and twenty five per million output. The small and fast tier of the same family costs 0.38 cents. At open-weight managed interface pricing typical of a model in the hundred billion parameter class, it costs 0.24 cents.

Set side by side, the three figures give two ratios.

Frontier to open weight is roughly 36 to 1. Frontier to the small tier of the same proprietary family is roughly 23 to 1. Now look at what that means in money rather than in ratios. The whole distance from frontier to open weight is 8.51 cents per request. The distance from frontier to the small tier of the same proprietary family is 8.37 cents, which is 98 percent of it. Moving down a tier inside whatever you already use captures almost the entire available saving, with no infrastructure change, no sovereignty change, no hiring and no requalification. Add caching, where a cached input costs on the order of 90 percent less than fresh input. Add batch processing at a consistent 50 percent discount. The remainder can then be cut again without changing suppliers.

Tier and cache selection is the largest cost lever available, and it belongs on the agenda before any change of model class.

The comparison that matters is self hosted against the open weight managed interface number above, a far harder bar, and not against frontier. Here is the published Canadian region retail pricing for accelerators suitable for open-weight inference, observed 12 September 2026, in United States dollars per accelerator hour. On demand: $8.59 to $14.75. One to five year committed capacity: $5.90 to $9.44. Spot or interruptible: $2.73 to $2.95.

Across those ranges, on demand runs roughly 2.9 to 5.4 times the spot rate and committed capacity roughly 2.0 to 3.5 times it. Identical silicon, identical country, and a spread of that size set by the purchase instrument. Anyone who cannot commit to multi year capacity or tolerate interruption is buying at the top of that range, and that alone can invert the crossover.

Then the utilization term. Self hosted cost is fixed by the hour, because the accelerator bills whether or not a request arrives. Hosted interface cost is fully variable per token. Effective cost per token equals the cost of one accelerator hour divided by the product of throughput in tokens per second, 3,600, and utilization. A tool running a fifty hour week is therefore capped at 29.8 percent utilization before any idle time within those hours. On a forty hour week the cap is 23.8 percent. At ten percent realized utilization, plausible for a single internal service, cost per token is ten times the fully loaded figure. The crossover volume moves by the same factor.

The crossover is a utilization threshold, not a volume threshold. Volume matters only insofar as it fills the accelerator.

One arithmetic warning. Published inference benchmarks report throughput on newer accelerators than the pricing above describes. That includes the MLCommons round of 1 April 2026, which added an open-weight model in the hundred and twenty billion parameter class. Dividing new generation throughput by previous generation prices understates cost per token, possibly by a factor of two to four.

We have not found a defensible published Canadian figure for any of these, so cost them yourself. Deployment and inference server tuning. Monitoring, plus token accounting and cost attribution. Requalification on every model update, which open weights do not receive automatically and hosted interfaces apply underneath you. Regression evaluation per version. Security patching across the inference server, driver stack and container images. And redundancy, which roughly doubles fixed cost while leaving throughput unchanged, and which a hosted interface prices in at zero marginal cost to the buyer.

Quebec’s provincial utility publishes a medium power rate. It applies to customers with at least 50 kilowatts of demand, effective 1 April 2026. The rate is 6.292 cents per kilowatt hour for the first 210,000 kilowatt hours and 4.666 cents thereafter, in Canadian dollars. A demand charge of $18.242 per kilowatt of billing demand comes on top. For contrast, Ontario Energy Board regulated rates effective 1 November 2025 put off peak at 9.8 cents and the first tier at 12.0 cents.

Now put that in the model. An eight accelerator node drawing roughly ten kilowatts including host and cooling overhead consumes about 7,300 kilowatt hours a month. That is well inside the first tier, so 6.292 cents applies. Energy comes to 7.87 cents per accelerator hour, plus 3.12 cents of demand charge. That is about eleven Canadian cents, or eight United States cents at roughly 1.37. A Canadian region on demand rate runs nearly fifteen United States dollars per accelerator hour. Against that, electricity is well under one percent of the cost. Against the cheapest committed rate it is about 1.4 percent.

Cheap hydroelectricity is decisive for siting a data centre. It is close to irrelevant to an enterprise deciding whether to self host, because that decision is dominated by capital recovery, purchase instrument and utilization. A cost model that leads with electricity is measuring the wrong thing.

Innovation, Science and Economic Development Canada’s sovereign artificial intelligence compute strategy carries an envelope of $2 billion over five years, Canadian. It spans fiscal years 2024 to 2025 through 2028 to 2029. Up to $700 million goes to mobilizing private sector investment. Up to $1 billion goes to public supercomputing infrastructure, and up to $300 million to a compute access fund. Its call for proposals on large scale sovereign data centres required planned capacity greater than 100 megawatts.

Read that threshold carefully. It is the scale at which Ottawa considers compute sovereign, orders of magnitude above any single enterprise deployment. Pricing a build as though it confers sovereignty is a category error.

Meanwhile, what Canadian organizations report as their barrier is not cost. Statistics Canada published an analytical article on artificial intelligence use by businesses for the second quarter of 2026 on 11 June 2026. It draws on the Canadian Survey on Business Conditions, fielded 1 April to 6 May 2026 with 9,251 responding businesses. The article found that 10.6 percent of all respondents cited the cost of using artificial intelligence as a barrier. That rose to 15.1 percent among businesses with 20 to 99 employees. Respondents cited cybersecurity or privacy concerns at 13.4 percent overall and at 30.0 percent among businesses with 100 or more employees.

Statistics Canada publishes no cost barrier figure for that largest size band, so you cannot set the two against each other directly. What we can say is that close to a third of the largest respondents named security and privacy, while cost never reached that level in any published cut. An awkward finding for a cost article, and the honest one.

We went looking for a credible, independent, dated, methodologically transparent total cost of ownership comparison of open-weight against proprietary deployment, and did not find one. Which means a supplier presenting you one is presenting an unvalidated internal model.

Build your own from the structure above, state the exchange rate and the utilization assumption, and put it in front of someone whose job is to disagree with it. Four inputs decide the answer: traffic shape, tier, utilization and purchase instrument. Everything else is rounding.

Previously: how to write a procurement-ready AI pilot statement of work. Next in this series: the business case for a six-week knowledge agent pilot.