Canada's Compute Access Fund Is Moving, But AI Isn't Waiting
Canada set out in 2024 to close the AI gap. The flagship $300-million Compute Access Fund (CAF) was meant to help SMEs cover the cost of building and running AI models. After consultations and a government change, applications opened in June 2025. Months later, some firms got rejection notices without reasons, others advanced, and no funds have been disbursed.
This is normal for government. It's slow for AI. Weekly breakthroughs don't pause while programs get stood up, evaluated, and announced. That gap is frustrating founders-and it signals a bigger strategic question for Ottawa.
What the CAF Set Out to Do
The CAF sits inside a broader $2-billion push to build AI data centres and compute infrastructure in Canada. The government has earmarked up to $240-million for Cohere and $42.5-million to the University of Toronto for research compute.
CAF itself targets SMEs. The intent: lower the cost of training models and running inference so more companies can participate. On paper, it's the right friction to remove.
Where Things Stand
Some applicants were rejected with no explanation. Others are still in evaluation for awards between $100,000 and $5-million. The process may slide into 2026. That lag is the core complaint.
"When you have a program that's a year-and-a-half in the making, you are a dinosaur by the point funds are released," said Moody Abdul of Klarify, whose CAF application was rejected. AI minister Evan Solomon counters that demand dwarfs supply: "There was absolutely no way that a $300-million compute access fund will be able to satisfy every single compute need."
The Strategic Debate: Models vs. Applications
CAF covers training new models and using existing ones. Some founders argue the real economic upside is on the application layer-software that solves clear problems-while foundational models and data centres get commoditized and drift into price wars.
Daniel Wigdor of Axl makes the point bluntly: building the grid is one thing; the value shows up in what runs on it. That mirrors the internet era, where software platforms outperformed carriers and hardware vendors. The ask: support the future Shopifys, not the next Nortels.
Operators echo this. Scott Stevenson at Spellbook says building proprietary general-purpose models is capital-heavy and rarely pays off for smaller firms. His team ships legal automation using existing models and has thousands of customers. Ryan Wilson at Walter Software adds that only a handful of well-funded players will win the foundational model race.
Counterpoints and Uncertainty
Not every CAF applicant proposed model training, notes Deloitte's Brendan Cooper. Some researchers also argue smaller, specialized models still matter. Vancouver's Variational AI is pursuing that path in drug discovery.
Evan Solomon says it's too early to judge the government's mix of bets-no CAF awards have been announced. He acknowledges the process has been slow, citing heavy application volume and diligence requirements.
Practical Risks for Government Teams
- Time-to-impact: If funds land in 2026, projects will be scoped on last year's assumptions.
- Over-index on supply: Compute alone doesn't create markets; demand and distribution do.
- Vendor concentration: Betting on generalized model training steers value to a few players.
- Commercialization gap: Research wins awards; application companies win customers and exports.
- Talent flight: Slow decisions push founders to relocate or re-incorporate elsewhere.
- Opacity tax: Rejections without reasons waste applicant time and erode trust.
What Government Can Do Next
- Publish timelines and criteria now: Set clear service levels for each stage; provide short, specific rejection feedback. Even two bullet points help teams adjust and reapply.
- Move to rolling, milestone-based disbursements: Offer smaller, fast first tranches within 30-45 days; tie follow-on funding to usage, revenue, or deployment milestones.
- Shift some support to demand-side incentives: Pilot SR&ED-style vouchers or tax credits for customers who buy from Canadian AI SMEs, letting the market pick winners. See the SR&ED program overview here.
- Stand up application-first institutes: Fund commercialization centres with KPIs tied to pilots, revenue, and exports-not just papers and patents.
- Simplify procurement for AI pilots: Create pre-approved catalogs, fixed-price pilot contracts under a threshold, and shared risk frameworks so departments can trial solutions in weeks.
- Offer shared compute credits with guardrails: Negotiate pooled credits across providers; require interoperability, cost transparency, and carbon reporting to avoid lock-in.
- Prioritize public-sector use cases: Fund projects that improve service delivery (claims, permitting, compliance) and can be replicated across jurisdictions.
- Report quarterly on portfolio mix: Publish the split between foundational model work and application-layer projects, along with outcomes.
How to Judge Success in 12 Months
- Median time from application to first disbursement.
- Number of SMEs shipping production features to paying customers.
- Public dollars leveraged by private capital and revenue (multiplier).
- Count of public-sector pilots deployed and scaled.
- Share of awards to application builders versus model training.
- Regional distribution and SME participation rate.
What Founders Are Asking For
Faster decisions. Clear criteria. Demand that helps close real deals. The message from builders such as Abdul, Stevenson, Wilson, and Wigdor is consistent: help them ship, not just train.
There are signs of movement. Budget changes signal intent to speed up and enhance SR&ED. If CAF pairs that with speed, transparency, and demand-side tools, Canada can still turn good infrastructure into great businesses.
For Teams That Need to Skill Up Fast
If your department is scoping pilots or evaluating vendors, a short skills sprint helps. Curated training by role can accelerate evaluations and procurement prep: AI courses by job.
Bottom Line
CAF can still deliver impact-but only if it moves faster and tilts toward applications that win customers, exports, and productivity. Build the grid, yes. Fund the factories running on it even more.
Your membership also unlocks:
