01 · The frame
The headline rate held. The base is collapsing under it.
Canada built the best non-dilutive R&D system in the world. The Scientific Research and Experimental Development program returns up to 35% of qualifying R&D spend as a cash refund, dispenses approximately $4.5B per year, and reaches roughly 20,000 Canadian businesses. For pre-revenue Canadian-controlled private corporations, it is often the single largest source of capital before Series A.
The November 2025 federal budget made the program more generous than it has been in over a decade. The enhanced expenditure limit went from $3M to $6M. The maximum refundable credit went from $1.05M to $2.1M per year. The taxable-capital phase-out range widened from $10M to $50M into $15M to $75M. Eligibility was extended to certain Canadian public corporations, and capital expenditures were restored.
Every one of those changes is directionally correct. None of them touches the line item that now dominates AI-native R&D budgets.
02 · The math
Three identical budgets. Three different outcomes.
| $5M R&D budget | Refund outcome |
|---|---|
| Canadian engineer salaries, in-house SR&ED | Up to $1.75M cash refund |
| Canadian taxable supplier performing SR&ED in Canada | Up to $1.40M cash refund, after the 80% rule |
| Owned hardware acquired after Dec 15, 2024, used substantially for SR&ED in Canada | Up to $1.75M, partially refundable |
| OpenAI, Anthropic, or Google API tokens for the same R&D | $0 |
The token line is not a rounding error and not an oversight. It is the direct, structural consequence of how the Income Tax Act enumerates eligible SR&ED expenditures. There is no category that catches it. In 2022, this gap touched a small number of teams running niche workloads. In 2026, it is the largest line item on most AI-native P&Ls.
03 · Six categories
Why tokens fall through every one.
Must be paid to employees who perform SR&ED in Canada. Tokens are not labour and are not consumed by the claimant's employees. Fails: not labour.
Physical inputs used up in the SR&ED process. Tokens are an output of a service, not a transformed input. Fails: not material.
Requires a taxable Canadian supplier, work performed in Canada, claimant ownership of the IP, and a contract scoped as SR&ED. None hold for US frontier-lab API calls. Fails: all four conditions.
Restricted to approved entities such as universities, colleges, and research institutes. US frontier labs are not on the approved list. Fails: not an approved entity.
Applies to depreciable property acquired and used substantially for SR&ED in Canada. A subscription invoice from a US vendor is consumption, not capital. Fails: service, not property.
The proxy method caps overhead at 55% of SR&ED salaries. As salary spend shrinks relative to token spend, the ceiling shrinks with it. Capped, or excluded.
The Income Tax Act has no category for an R&D input that is variable, metered, served from outside Canada, and not owned by either party. Token consumption is the first major R&D input in 40 years that does not fit the original design assumptions of the program.
04 · The trend
Token spend is the structural form R&D is taking.
Year over year, 2023 to 2024. Stanford HAI 2025 AI Index.
Annual, 2024 baseline.
$24B to $150B enterprise AI, 2024 to 2030.
$200 to $1,200 per month, 2025 to 2027.
The trajectory inside R&D itself is steeper than the enterprise average. Public usage data shows 30M to 50M tokens per engineer per day in active agentic coding workflows. At a $15 per million blended frontier price, a single engineer carries roughly $90K of token spend annually before any product workload is added. A 30-engineer Series A runs $2.7M in pure engineering token cost. Add product inference at the typical 5x to 10x ratio and the company is at $13M to $27M in annual token cost by Series B. None of this existed as a P&L line in 2022.
05 · The Canadian math
$4.1B in tokens. $4.5B in SR&ED.
Apply the trend to the Canadian tech workforce. Roughly 2.3M tech workers nationally. Conservative scenario: 10% become AI-native by 2027. That is 230,000 workers at $1,500 per month average AI tooling, or approximately $4.1B per year in tokens flowing out of the country. By 2027 the ecosystem is on track to spend, on AI services alone, an amount roughly equivalent to the entire annual SR&ED program. Zero of it qualifies.
Every Canadian AI-native company is silently absorbing a 25 to 35 point increase in net R&D cost compared to a peer doing the same work three years ago. The realized credit, as a percentage of total R&D spend, is dropping toward zero.
Top AI talent gravitates toward organizations providing modern tooling. Companies that throttle token spend to preserve their SR&ED ratio are the same companies losing their best people.
The $6M ceiling went up at exactly the moment the eligible base started collapsing under it. The reform is unintentionally rewarding old-shape companies and penalizing new-shape ones.
The SR&ED ratio is a primary input to Canadian VC valuation models. As the realized credit drops, the Canadian discount on AI-native companies widens against US comparables.
06 · The adjustment
Three principles, ranked by urgency.
Treat eligible AI inference and training services analogously to consumed materials. Require documentation tied to specific SR&ED projects via prompt logs and API trace IDs. Cap eligibility per project and audit on the same cycle as overhead.
Full eligibility for tokens served on Canadian-resident infrastructure, a reduced rate or cap for foreign-served tokens. Aligns the demand side with the supply-side SCIP program.
Recognize that token spend is uniquely metered, traceable, and attributable. Eligibility determined by the SR&ED purpose of the inference, not the static category of the invoice.
07 · The stake
The fix needs the next budget cycle, not the one after.
The 2026 reforms are correctly read as a vote of confidence in Canadian R&D. They are also, as written, a bet on a 2010-vintage cost structure. Canada is investing more than $4B in sovereign AI capacity across multiple programs, all on the supply side. The demand side has not been re-architected. Canada built the best non-dilutive R&D system in the world. The decade ahead runs on tokens. The system needs to catch up before the math forces every AI-native founder to incorporate somewhere else.