Dartmouth Advisory Partners

State of Canadian compute, November 2026

Data to 5 November 2026. Dartmouth Advisory Partners.

The month in one screen

GPU hours cost 15% to 20% more in Canada than in their cheapest market, and every price shown here comes from a foreign-parent seller. On 5 November, Budget 2026 promised Canadian AI companies access to 500 MW of sovereign compute. Of the six capacity statements we track, one is operating: 20 MW.

The month in one screen
QuestionAnswerWhat it counts
Does a GPU hour cost more?15% to 20% moreA100, H100 and GB200 on demand, each against its own cheapest market
Do model tokens cost more?9 of 29 models6.7% to 16.7% more than the same seller elsewhere. The other 20 match
Who controls it?7 of 10 venues, 0 of 67 pricesVenues with a Canadian parent, against prices shown here from a Canadian-parent seller
What is running today?1 of 6 capacity statements20 MW at SPUR Innovation Centre, Waterloo campus. The rest are secured or planned

The register holds 69 lines hosted in Canada: 6 GPU, 61 token and 2 power. Market data runs to 5 November 2026. Policy runs to 6 November 2026.

Does Canada cost more?

Yes, for all three GPUs priced on demand in Canada. The gap runs from 15.0% to 20.0%, and the same seller sits on both sides of every pair.

Does Canada cost more?
GPULowest in Canada (US$ per GPU-hour)Cheapest market (US$ per GPU-hour)Canada costs
A100 40 GiB$3.16, AWS$2.74, AWS, US East (Ohio)+15.0%
H100 80 GiB$7.91, AWS$6.88, AWS, US East (N. Virginia)+15.0%
GB200 192 GiB$32.45, Azure$27.04, Azure, East US+20.0%

Gaps come from the unrounded list prices, one GPU at a time. They are never averaged. H200, B200 and B300 are not listed on demand in Canada, so they have no pair. B300 ranges from $17.80 to $35.14 across the console.

What do model tokens cost?

Most models cost the same in Canada: 20 of 29 match their seller's lowest price elsewhere. The other 9 cost 6.7% to 16.7% more.

All 17 Anthropic models match. Every Mistral, Meta and Amazon model costs more.

One square per model: 9 cost more in Canada (blue), 20 match the seller's lowest price elsewhere (grey).

All 17 Anthropic models match. Every Mistral, Meta and Amazon model costs more.Meta: 2 cost more in Canada, 0 match. +15.0% to +16.7%. Mistral: 3 cost more in Canada, 0 match. +13.3% to +15.7%. OpenAI: 2 cost more in Canada, 1 match. +10.0%. Amazon: 2 cost more in Canada, 0 match. +6.7% to +8.0%. Anthropic: 0 cost more in Canada, 17 match. Same as elsewhere. Moonshot AI: 0 cost more in Canada, 1 match. Same as elsewhere. xAI: 0 cost more in Canada, 1 match. Same as elsewhere.Meta+15.0% to +16.7%Mistral+13.3% to +15.7%OpenAI+10.0%Amazon+6.7% to +8.0%AnthropicSame as elsewhereMoonshot AISame as elsewherexAISame as elsewhere
DAP compute console, Canada view, 5 November 2026. 29 models with an input and an output price. Gaps are derived against the same seller's lowest price elsewhere.

AWS sells on both sides of this split, so the seller alone does not explain it. Of the 61 token lines, 58 pair into these 29 models. The other 3 have no input and output pair.

Who controls it?

Canadian parents own most venues on the register, yet foreign-parent sellers set every price shown here. That is 7 of 10 venues against 0 of 67 prices.

Who controls it?
ControlVenuesPrices shown here
Canadian parent, serves in CanadaISAIC, TELUS Sovereign AI Factory, Bell AI Fabric, Hypertec Cloud0
Canadian parent, mixed railsSPUR Compute, Cohere, Augure0
Foreign parent, hosted in CanadaAWS, Microsoft Azure, OVHcloud67

SPUR Compute, a Canadian-parent venue, publishes prices. None is shown here yet, because a price published in two places is shown only once the two agree. The gap therefore measures published, comparable prices. It does not measure what Canadian venues charge under contract.

What is being built?

Most of what companies have stated is not running yet. Of 6 capacity statements across 4 sites, 1 is operating, 3 are secured and 2 are planned.

Of six stated figures, one is operating: 20 MW at Waterloo.

Each company's own statement, in MW. Blue is operating, grey secured, light grey planned. Do not add these bars.

Of six stated figures, one is operating: 20 MW at Waterloo.SPUR Waterloo campus, Operating today: 20 MW, operating. Bell AI Fabric, Sherwood, SK, New data centre: 300 MW, secured. TELUS BC cluster, By 2032: 150 MW, planned. SPUR Canadian portfolio, Available for contract: 100 MW, planned. TELUS BC cluster, Initial power secured from BC Hydro: 85 MW, secured. SPUR Waterloo campus, Committed expansion: 20 MW, secured. Reference line: 500 MW Budget 2026 goal. Reference line: 850 MW proposed by 2030.850 MW proposed by 2030500 MW Budget 2026 goalSPUR Waterloo campusOperating today20 MWBell AI Fabric, Sherwood, SKNew data centre300 MWTELUS BC clusterBy 2032150 MWSPUR Canadian portfolioAvailable for contract100 MWTELUS BC clusterInitial power secured from BC Hydro85 MWSPUR Waterloo campusCommitted expansion20 MW0300600900 MW
Company statements on the DAP register, March to October 2026. Federal figures from Budget 2026 and AI for All.

The figures can overlap, so they are never summed. The register covers these 4 sites, not all of Canada. The federal lines are goals: AI for All also names a path to 2.3 GW, which is off this chart.

What Ottawa has committed

Ottawa's newest tool backs demand, not only hardware. Budget 2026 says much new capacity is being secured by large technology companies on long contracts, leaving Canadian companies short at home. Its answer is 500 MW through backstops and offtake agreements.

Amounts are Canadian dollars. Newest first.

What Ottawa has committed
DateMeasureAmountWhere it stands
5 Nov 2026Budget 2026: 500 MW of sovereign AI compute for Canadian AI companiesNo dollar figure in the chapterAnnounced, through new financial tools
5 Nov 2026Budget 2026: AI for All$2.7 billion initial, cash basisFiscal table shows $2,038 million from 2026-27 to 2030-31. The chapter does not reconcile the two
29 Oct 2026Cohere stakeUp to about $1.6 billion in equity, in stagesCohere commits at least $1.7 billion to additional computing infrastructure in Canada by 2030
4 Jun 2026AI for All: Compute Access Fund expansion$700 million more, for SMEsStrategy commitment
4 Jun 2026AI for All: private partnerships850 MW proposed by 2030, scaling to 2.3 GW"Being finalized". Proposed, not contracted
11 May 2026TELUS BC projectNot stated"No funding has yet been committed or distributed"
15 Apr 2026AI SCIP: public supercomputer buildUp to $890 million over seven yearsApplications closed 1 June. We found no recipient announced as of 6 November
2024Sovereign AI Compute Strategy$2 billion over five yearsAI for All says delivery continues

AI SCIP spells out what sovereign means in a federal program. Core compute and storage must be owned or contractually controlled by Canadian entities, for the full life of the facility. For-profit companies can only join as consortium partners.

Dates to watch

The next signal is who wins AI SCIP. Winners must offer significant service within 18 months of signing, so that name also sets the real clock.

The nearest decisions have no date. Stated milestones start in 2029.

Blue has no published date. Grey is a date stated by Ottawa or the company.

The nearest decisions have no date. Stated milestones start in 2029.No published date: AI SCIP recipient named; 500 MW backstops and offtake terms; Partnership agreements finalized. Stated dates: TELUS: 150 West Georgia online, 2029; Partnerships: 850 MW proposed, 2030; Cohere: $1.7 billion compute, 1,000 jobs, 2030; Public AI supercomputer, 2031; TELUS: 150 MW, over 60,000 GPUs, 2032.No date2026202820302032AI SCIP recipient named500 MW backstops and offtake termsPartnership agreements finalizedTELUS: 150 West Georgia online2029Partnerships: 850 MW proposed2030Cohere: $1.7 billion compute, 1,000 jobs2030Public AI supercomputer2031TELUS: 150 MW, over 60,000 GPUs2032
Budget 2026, AI for All, Torys on AI SCIP, TELUS. Dates as each source states them.

Method and disclosure

  • Prices. Each line is one seller's own number, read by machine on the date shown. Nothing is averaged or summed across sellers.
  • GPU gaps. Lowest on-demand list price in Canada against the lowest for the same GPU and memory in any market, in US dollars per GPU-hour.
  • Token gaps. The same seller's lowest price for the same model and direction in its other regions.
  • Control. Derived by a fixed rule from each venue's cited parent and place of service. A line takes its seller's control.
  • Capacity. Shown as each company states it, in its own words. Never summed.
  • Coverage. The register is not a census of Canadian compute. A venue with no published price will not appear in the price counts.

Dartmouth Advisory Partners has no commercial relationship with any vendor covered here. This report is not a DAP price, an index or investment advice. Government measures are reported as published, and DAP takes no position on them.

Sources