SR&ED and Canadian AI Infrastructure: What Counts as Qualifying Expenditure

Canada's Scientific Research and Experimental Development (SR&ED) tax incentive program is one of the most generous R&D programs in the OECD. For technology companies, it offers a refundable 35% investment tax credit (for Canadian-controlled private corporations) on qualifying expenditures. For companies building AI infrastructure, the SR&ED rules create real opportunities — but also a distinction that founders frequently miss, and that CRA auditors frequently scrutinise.

This post explains what qualifies as SR&ED work in the context of AI infrastructure development, what does not, and how to structure your claim to reflect the line accurately.

The Three SR&ED Criteria

Work qualifies as SR&ED under the Income Tax Act if it meets three criteria simultaneously:

• Scientific or technological uncertainty: The outcome of the work was not known or determinable in advance by a competent professional in the field.
• Systematic investigation: The work followed a defined hypothesis, methodology, and evaluation process — not ad hoc troubleshooting.
• Advancement: The work attempted to advance scientific or technological knowledge, not just apply existing knowledge.

The first criterion — technological uncertainty — is where most AI infrastructure claims succeed or fail. CRA defines technological uncertainty as uncertainty about whether a goal can be achieved or how to achieve it, where the answer is not readily deducible from existing knowledge in the field.

What Qualifies: Building the Sovereign Inference Stack

The development work involved in building Sovereign AI Gateway's infrastructure illustrates what qualifies. Consider the following problems our engineering team faced:

Low-latency inference under strict network isolation

Modern LLM inference stacks (vLLM, TensorRT-LLM) are designed assuming outbound network access for telemetry, model updates, and distributed coordination. Achieving competitive inference latency while operating under complete outbound egress lockdown — where zero packets leave the colo boundary except to specific allowlisted destinations — required solving problems that had no published solutions at the time. Adapting these systems to a fully isolated environment involved technological uncertainty about whether the performance targets were achievable at all.

Cryptographic attestation at inference time

Designing a per-request sovereignty attestation system — one where the signed receipt is produced at inference time with negligible latency overhead — involved uncertainty about how to structure the signing flow to avoid becoming a bottleneck in high-throughput scenarios. The solution required novel approaches to key management and signing pipeline design.

Audit logging without content retention

PHIPA's requirements and our customers' sovereignty guarantees require us to maintain detailed audit logs without retaining the content of any request or response. Designing a logging architecture that satisfies regulators' audit requirements while cryptographically guaranteeing content non-retention involved genuine technical problems with no obvious prior-art solution.

What Does Not Qualify: Running the Commercial API

Once the sovereign inference stack is built and operational, the work of running it as a commercial service does not qualify as SR&ED. This is the line that matters:

• Operating a production API on infrastructure that already works — not SR&ED.
• Routine scaling of proven infrastructure — not SR&ED.
• Customer onboarding, support, and account management — not SR&ED.
• Integrating existing open-source components without modification — generally not SR&ED.
• Market research, business development, and compliance documentation — not SR&ED.

CRA is consistent on this point: commercialisation of a completed technology is not R&D, even if the technology itself was developed through qualifying SR&ED work. The claim must be scoped to the development phase, with contemporaneous records documenting when that phase ended and the commercial phase began.

Claim Structuring: What CRA Looks For

SR&ED claims in the AI infrastructure space attract CRA scrutiny because the boundary between "building novel technology" and "applying existing frameworks" can be contested. Strong claims share several characteristics:

• Clear technological hypothesis: Each SR&ED project should have a documented hypothesis — a specific technical problem statement, a proposed approach, and success criteria.
• Contemporaneous records: Lab notebooks, design documents, commit logs, architecture decision records, and test results dated during the period of uncertainty. Records created at claim time carry significantly less weight than records created during the work.
• Defined endpoints: The claim should identify when the work resolved the technological uncertainty — not just "we kept improving the system."
• Competent professional attestation: A CTO or senior engineer who can explain why the problem was uncertain at the time and what competing approaches were evaluated.

Interaction with Other Programs

SR&ED claims for AI infrastructure may interact with other Canadian programs worth noting:

• NRC IRAP: Industrial Research Assistance Program funding for SMEs can be combined with SR&ED, but IRAP contributions typically reduce SR&ED-eligible expenditures by the amount received. Structure funding applications in sequence or with awareness of the offset rules.
• SDTC: Sustainable Development Technology Canada funding (now under ISED management) may be relevant if the AI infrastructure work has an environmental efficiency dimension — for example, optimising inference efficiency to reduce per-token energy consumption.
• Ontario Innovative Business Tax Credit: Ontario's 8% provincial credit stacks on top of the federal SR&ED ITC for work performed in Ontario, available to CCPCs that qualify federally.

If you are building Canadian AI infrastructure and believe your development work may qualify, the key steps are: document the technological uncertainties as you encounter them, maintain records of the systematic investigation, and engage a qualified SR&ED consultant to review your work before filing. The credit is substantial — but only for work that genuinely involved resolved technological uncertainty, documented in a way that survives a CRA technical review.