AI Compute Access Fund 2026: The Complete Guide for Canadian AI Builders

If you are training, fine-tuning, or running inference on AI models out of Canada, the federal government will now reimburse a large chunk of your compute bill — provided you spend that compute on Canadian soil. The AI Compute Access Fund is the access pillar of Canada's Sovereign AI Compute Strategy and is the most directly useful AI-specific program for builders ISED has launched in years. It is also a program whose parameters are still settling, where the definition of "domestic compute" drives most of your eligibility math, and where stacking with SR&ED and Scale AI requires some genuine thought. This guide walks through how the fund actually works, what it pays for, where applications break, and how to slot it into a Canadian AI company's broader funding stack.

What the AI Compute Access Fund is

The AI Compute Access Fund is a federal contribution program administered by Innovation, Science and Economic Development Canada (ISED). It pays back a portion of what a Canadian AI builder spends on cloud-based GPU and AI compute services — the picks and shovels of modern AI work — with a strong bias toward keeping those dollars inside Canadian-hosted infrastructure. In ISED's own framing, the fund exists because consultations identified "the high cost of compute resources and the limited availability of domestic capacity" as the most acute barrier to Canadian AI builders staying and scaling here.

That problem statement is worth pausing on. Most of Canada's AI talent is here. Most of Canada's AI startups are incorporated here. But for the last several years, a meaningful share of the actual training and inference workloads ran on US-hosted hyperscaler GPU capacity — because that's where the H100s, the A100s, and now the newer accelerators were physically available, and because pricing on Canadian-hosted equivalents was either thin or non-existent. The Sovereign AI Compute Strategy, announced in 2024, is the federal response to that imbalance. The Access Fund is the demand-side lever: subsidize Canadian AI builders' bills when they spend on Canadian-hosted compute, and the domestic supply will follow.

Mechanically, the fund is a reimbursement-style contribution. You apply with a project plan that describes the AI work, the compute services you intend to consume, the providers you intend to use, and the business case for the work. If approved, you enter a contribution agreement that defines the eligible spend and the cost-share rate. You then incur the compute costs against your AI workload, submit claims, and receive a non-repayable contribution back from ISED at the agreed share.

How it fits into the Sovereign AI Compute Strategy

To use the Access Fund well, it helps to understand the broader strategy it sits inside. Canada's Sovereign AI Compute Strategy is a roughly $2 billion package announced in 2024, structured around three pillars that work together:

The strategy is deliberately designed so that the supply pillars and the demand pillar pull in the same direction: data centres get built, public capacity comes online, and Canadian AI companies have a cost-shared incentive to consume that capacity. As a builder, you don't need to engage with the supply pillars directly — you'll just see them show up over time as more Canadian providers list compute on procurement-friendly terms. Your direct interaction is with Pillar 3.

The cost-share rates and the $100K–$5M range

This is where most of the planning work happens. The AI Compute Access Fund publishes two cost-share rates, and which one applies to your project depends entirely on where the compute is hosted:

That 16-point spread between the two rates is not accidental. It's the federal government putting a number on how much it values keeping AI compute spend inside Canadian infrastructure. For a project burning $1M in GPU costs, choosing a Canadian provider can mean roughly $660K reimbursed versus roughly $500K — a $160K differential that, in many cases, will outweigh modest price differences between Canadian and non-Canadian providers on a like-for-like instance.

Project sizes are bounded on both ends. The fund supports projects from $100,000 up to $5 million in eligible compute spend, typically structured over up to three years. The lower bound exists because the program isn't optimized for one-off training runs — the application overhead doesn't make sense below roughly $100K of planned spend. The upper bound caps any single project's draw on the $300M envelope so that the fund can support a portfolio of builders rather than a handful of giant training runs.

One practical note on the math: the cost-share is applied to eligible compute costs, not necessarily your full invoice. Eligible costs are the AI-relevant portions of cloud spend — GPU instances, accelerator time, AI-specific managed services, and the storage and networking directly tied to those workloads. General-purpose cloud spend, internal IT infrastructure, and non-AI services aren't part of the eligible base. Most builders will end up apportioning their cloud bill into AI-workload and non-AI-workload buckets when scoping the application.

What "domestic compute" means and who qualifies

This is the part of the program that drives more application work than any other single question. "Canadian cloud-based AI compute" is a deceptively simple phrase that covers several distinct provider archetypes:

• Canadian-incorporated cloud and GPU providers operating data centres physically located in Canada (Telus, Bell, smaller Canadian GPU operators, and the wave of new entrants that the AI Compute Challenge is funding).
• Canadian sovereign-region offerings from international hyperscalers — specifically the Canadian regions of providers like AWS (ca-central-1), Microsoft Azure (Canada Central/East), Google Cloud (northamerica-northeast1/2), and Oracle Cloud (Canada Central). The eligibility test is whether the compute is physically located and contractually delivered from a Canadian region, not the corporate nationality of the parent provider.
• Canadian academic and consortium compute available on commercial terms to companies, where applicable to the project.

What is not domestic compute for the purposes of the higher cost-share rate: training runs you execute on US, European, or Asian regions of any provider — even if the provider has a Canadian subsidiary, even if the data is Canadian, and even if you're paying the bill from a Canadian entity. The location of the GPUs is the test.

Practically, this means a Canadian AI startup that has been running training on AWS's us-east-1 because that's where its team set up the account is going to want to assess whether ca-central-1 can carry the same workload before applying. In many cases it can, with some additional configuration. In a smaller number of cases — very large training runs requiring specific instance types or contiguous capacity not currently available in Canadian regions — the workload genuinely needs to stay south of the border, and the 50% non-Canadian rate applies. The fund anticipates both scenarios.

Eligibility — Canadian AI builders

The applicant-side eligibility test is fairly clean compared to some federal programs. Based on the most recent ISED intake materials, an eligible applicant is:

• A Canadian-registered for-profit company — incorporated federally or provincially in Canada
• With fewer than 500 full-time equivalent (FTE) employees — the SME definition the federal innovation programs generally use
• Either revenue-generating or able to demonstrate Series A (or later) financing — the program wants companies past the earliest validation stage
• Maintaining an R&D team based in Canada — the AI engineering work needs to be happening here
• With credible commercialization plans for the AI product or capability being developed

The "AI builder" framing is important. The fund is for companies that develop, train, fine-tune, or operationally deploy their own AI models — not for businesses that consume third-party AI APIs as an end user. If your compute spend is mostly inference calls to OpenAI or Anthropic from your application backend, you're not the user the program is designed for. If you are training proprietary models, fine-tuning open-weights models on your own data, or running large-scale inference on models you build, you are.

How to apply (the ISED portal)

The application flow follows the standard ISED contribution-program shape. The mechanics matter less than the timing, so what follows is the structure most builders should expect; check the live ISED page for exact deadlines, document templates, and account setup links before starting.

A note on intake windows: the most recent published application window closed in mid-2025. The program is multi-year, so subsequent intake windows have followed and more are expected. If the portal currently shows "closed pending assessment," that is not the end of the program — it's a normal between-intake state. ISED announces upcoming windows on the Sovereign AI Compute Strategy page and through the regular federal innovation funding channels.

Stacking with SR&ED and Scale AI

The Access Fund pays for compute. SR&ED pays for engineering labour. Scale AI pays for industry-led AI commercialization projects. These three programs cover different parts of the same AI company's spending and, with reasonable care, can stack on the same underlying initiative.

AI Compute Access Fund + SR&ED

This is the most useful combination, and the one most builders should plan for. SR&ED is a self-assessed credit on the engineering work itself — the time your team spends on scientific and technological advancement, scoped through the S/THERI framework (Scientific/Technological uncertainty, Hypotheses, Experimentation, Results, Iteration) and documented per CRA IC 2012-02. The Access Fund subsidizes the compute those engineers are running. Different cost categories, different programs, different filing mechanics.

Where it gets interesting is the overlap: if the Access Fund is reimbursing two-thirds of an eligible compute cost that also would have been a SR&ED-eligible expenditure (compute used for the experimental work itself), CRA treats the federal contribution as government assistance that reduces the SR&ED claim on that specific cost. The same dollar can't be funded twice. In practice, this is usually a small adjustment because the bulk of most SR&ED claims is labour, not compute — but it has to be tracked and disclosed on the T661.

Two operating principles:

• Separate the line items. SR&ED claims engineering labour. The Access Fund claims compute. Where compute is also being claimed as a SR&ED material consumed, net the Access Fund contribution off the SR&ED-eligible amount.
• Disclose both programs in both filings. The SR&ED claim discloses the Access Fund contribution as government assistance. The Access Fund claim discloses SR&ED filings against the same project. Silence in either direction is a faster path to a reassessment than the actual stacking math.

AI Compute Access Fund + Scale AI

Scale AI funds consortium-led AI commercialization projects — an AI vendor and an industry Solution Adopter working together to deploy AI into a real production setting. The Access Fund funds the compute side of a Canadian AI builder's broader work. These programs aren't substitutes; in some cases they're complementary on the same engagement.

A common pattern: a Canadian AI vendor has a Scale AI project deploying its model into a manufacturing customer's operations. The Scale AI project covers the engineering and integration work. Separately, that same vendor has continuing training and inference workloads — serving the deployed model, running ongoing retraining cycles, supporting other customers — and the Access Fund covers a share of the underlying compute bill on those workloads. Two programs, two cost categories, one company.

The stacking rule that matters is the standard federal one: government assistance against the same eligible expense can't exceed defined caps (typically 75% of that cost across all sources). At the Access Fund's cost-share rates and Scale AI's roughly one-third rate, you would have to deliberately structure a project to bump into that cap on the same cost line. The risk is real but manageable with reasonable budget hygiene.

Common reasons applications won't fit

The Access Fund is younger than CanExport or SR&ED so the pattern of common failures is still emerging, but the early signal from ISED's published criteria and from applications we've seen reviewed is consistent:

• Compute scope below the floor. A project with $40K of planned compute spend isn't a fit, even if the AI work is excellent. The application overhead doesn't make economic sense for either party. Smaller compute needs sit better inside a SR&ED claim (where the materials get rolled in with labour) than a standalone Access Fund application.
• "Domestic" compute that isn't. Budgets built on US-region pricing that get re-labelled as Canadian without actually changing the workload location. Reviewers can read invoices and can ask for regional confirmation from the provider. This shows up as variance during claim review and erodes trust on the contribution agreement.
• API-consumer business model. A SaaS company whose AI is mostly third-party API calls is not the target user. The fund subsidizes compute spent by AI builders on workloads they own and run, not subscription costs for hosted AI services from US vendors.
• No commercialization story. "We will train a model" without a credible answer to "and what will you do with it" is a research proposal. The Access Fund is for builders with a commercial path; pure research belongs at NSERC, Mitacs, or in an academic-industry partnership program.
• R&D team not in Canada. Multinationals trying to claim the fund for workloads supporting non-Canadian engineering teams. The R&D-in-Canada requirement is a real test, not a rounding error.
• Underbaked compute plan. A budget that lists "cloud compute — $400K" with no breakdown across providers, instance types, training and inference, or workload phases. Reviewers want to see that the applicant has actually planned the compute, not just estimated a line item.
• Working capital mismatch. Reimbursement-style contributions require the recipient to carry the spend until the claim is paid. A company that can't fund several months of compute bills before the first reimbursement lands will struggle mid-project.

Final thoughts

The AI Compute Access Fund is the most direct federal subsidy that an operating Canadian AI builder has access to in 2026. It pays for the line item — cloud compute — that, after engineering labour, is usually the second-largest cost on the income statement of any serious AI company. At two-thirds cost-share on Canadian-hosted compute, the math is straightforwardly attractive for any company spending $100K or more annually on training, fine-tuning, or inference workloads.

The two things to keep in mind: first, the program is one piece of a broader Sovereign AI Compute Strategy, and program parameters — intake schedules, eligibility thresholds, cost-share rates — are likely to keep being refined as ISED works through the strategy's multi-year arc. Treat published figures as the current best information and verify against ISED before locking budget against them. Second, the domestic-compute requirement is doing real work in the program design — the 16-point spread between Canadian and non-Canadian rates is meaningful enough that builders should genuinely assess whether Canadian-region capacity can carry the workload before defaulting to US regions out of habit.

For most Canadian AI companies the right funding stack in 2026 looks something like this: SR&ED on the engineering labour spent on genuinely uncertain technical work, the AI Compute Access Fund on the compute bill that powers that work, Scale AI on the commercialization projects where there's a customer co-investing in deployment, and IRAP on the late-stage product-readiness work that doesn't fit cleanly into any of the above. Use our Grant Finder to compare the Access Fund against SR&ED, Scale AI, and other federal programs before committing application time. Each program has a specific shape and a specific kind of work it funds well; lining the right program up to the right cost category is most of the planning job.