Diary reconciliation — October 11, 2026: The original checks were April 6, followed by April 14 and April 16 updates. The summary now reflects those updates. A provider catalog, account quota and successful launch are different evidence; this diary is not a current Canadian GPU inventory or an exhaustive provider survey. The three-weeks-later follow-up is part of this record. Google’s accelerator documentation describes A3 Edge as inference-optimized; that is not evidence that its H100s cannot perform training.
Availability and pricing in this post reflect what I found as of April 6, 2026. Cloud providers update their infrastructure regularly, and I hope some of this changes soon. If you find something different, I'd genuinely love to hear about it.
This is a series of me publicly documenting my journey. The next post is me documenting the Three Weeks Later.
I chose Canada.
That might sound like a strange way to start a post about cloud computing, but it matters.
I'm not Canadian by birth. I came here, built my life here, built my company here. This isn't a temporary arrangement for me. I'm committed to this country for the long term.
So when I decided to build AI — real AI, not a wrapper on someone else's API — I wanted to do it here. Canadian company. Canadian data. Canadian compute.
I assumed the infrastructure would be there. Canada has a federal AI strategy. A compute access fund. Sovereignty frameworks.
We host MILA, one of the best AI research labs in the world. The talent is here. The policy ambition is here.
Then I tried to actually rent a GPU.
What I Was Looking For
My company needs GPU compute for AI model training.
Not a side project. Real training runs on current-generation chips — H100s, the same hardware that every serious AI lab in the world is using right now.
The requirements were simple. I wanted GPUs physically located in Canada. Current generation. Available on demand. Clean invoicing.
Shouldn't be hard, right? This is 2026. Canada is supposed to be an AI country.
I spent the better part of two weeks checking every cloud provider I could find with Canadian data centers. Google Cloud. AWS. DigitalOcean. OVHcloud.
A non-profit in Alberta. I called sales teams. Dug through documentation pages. Spun up test instances. Read fine print.
What I found made me angry. Then worried.
Google Cloud Montreal: 2017 Called
Google Cloud has two Canadian regions. Montreal and Toronto.
I started with Montreal, their primary Canadian region. The one you'd expect to have the full range of services.
Here's what's actually available for GPU compute in Montreal:
The NVIDIA P4. Released in 2017.
And the NVIDIA T4. Released in 2018.
That's it.
No H100. No A100. No L4. No V100. Nothing released in the last four years. No TPUs. Not a single piece of current-generation AI training hardware.
I checked this multiple times because I genuinely didn't believe it.
I went to Google's own GPU locations page, filtered for northamerica-northeast1, and stared at the screen.
P4 and T4. In 2026.
Update — April 14, 2026:
After this post published on April 7, Google added L4 (G2) instances to Montreal zones b and c.
A small improvement for the configurations I was evaluating. Calling L4 inference-oriented does not establish that it cannot perform training.
Toronto already had L4. My April finding was that I had not secured the training configuration and capacity I needed in either Canadian Google Cloud region; Toronto’s listed A3 Edge H100s mean “no H100” is too broad.
I'm noting this because it matters: this landscape is moving faster than static blog posts can track, which is exactly why anyone making infrastructure decisions should verify current availability before committing budget.

For context, a single zone in Iowa — just one of dozens of US zones — has H200s, H100s in multiple configurations, A100s, L4s, and everything in between.
Iowa alone has more GPU variety than all of Canada combined on Google Cloud.

And here's the part that really got under my skin.
If you go to third-party pricing comparison tools — the sites founders use to make infrastructure decisions — they list H100 pricing for Montreal.
The SKU exists in Google's system. The price shows up. It looks like you can rent an H100 in Montreal for $9.80 an hour.
You can't.
The pricing page and the availability page tell two completely different stories. One is marketing. The other is reality.
And if you don't know to check both, you'll plan your entire infrastructure around hardware that doesn't exist where you think it does.
Google Cloud Toronto: Close, But Not Quite
Toronto, Google's second Canadian region, is slightly better. It has L4 GPUs and a limited H100 configuration called "A3 Edge."
But A3 Edge is specifically designed for inference and serving — running already-trained models. Google's own documentation says this.

Whether Toronto’s A3 Edge configuration suits a training workload depends on its requirements and actual capacity; inference optimization alone does not mean training is impossible.
So across both Canadian regions, I had not found a launchable configuration suitable for my intended training run. That is a result for my search and account, not a technical claim of zero training options.
AWS Canada: Better, But You'll Still End Up in the US
AWS's Canadian region is better than Google's on paper. I'll give them that.
The ca-central-1 region does list current-generation training SKUs — P5 (H100) and P4d (A100) both appear in the EC2 instance types catalog. That's more than Google Cloud Montreal can say.
But listed isn't the same as launchable.
Update — April 14, 2026:
I went back and checked my own AWS account to see what I could actually run in ca-central-1. The answer is nothing.
Every GPU instance family — P (H100/A100), G (L4/A10G/T4), Inf (Inferentia) — has a default service quota of zero vCPUs on my account in Canada.
I can't launch a p5.48xlarge. I can't launch a p4d.24xlarge. I can't even launch a g4dn.xlarge with a T4 from 2018.
Not because the hardware isn't there in AWS's catalog — it is. Because my account isn't approved to run any of it in this region without submitting a quota-increase request and getting a human at AWS to say yes.

For P5 in non-US regions in 2026, that human approval is non-trivial. A common outcome is a polite redirect to US regions, where AWS has actual capacity to allocate.
Another common outcome is a capacity waitlist even after approval, because a quota is a ceiling, not inventory.
So the honest state of AWS Canada is this: the SKUs exist, but my account needed a quota increase before I could attempt a launch. That did not establish inventory or prove a structural regional bias.
For my account at that point, US regions were a fallback I was considering; this does not establish the route every AWS customer would need. Your data leaves Canada. Your workloads run on American soil.
For a company trying to build sovereign AI infrastructure, "your best option is to submit a quota request and hope AWS lets you run it in Canada, or fall back to the US" is not an answer.
OVHcloud: The One That Almost Fooled Us
OVHcloud is a French company with a data center in Beauharnois, Quebec. On paper, a great Canadian option.
Established company. Competitive pricing. H100 GPUs listed on their website.
I was ready to allocate budget.
Then I checked which GPUs are actually in their Canadian data center.
V100 and V100S. Previous-generation hardware from 2017 and 2018.
The H100s on their pricing page? Gravelines, France.
Not Canada. France.
If I hadn't dug into region-specific availability — if I'd just trusted the headline pricing like most busy founders would — we would have committed real money to a provider that cannot deliver what we need in the country we need it in.
This keeps happening. The marketing says one thing. The infrastructure says another. And nobody's talking about the gap.
ISAIC: Hope, But Not Yet
ISAIC is a non-profit in Edmonton, connected to the University of Alberta. They claim H100 access at $2.50 CAD per hour.
I wanted this to work more than any other option on this list.
A Canadian non-profit offering affordable AI compute to Canadian companies?
That's exactly what should exist. That's the kind of infrastructure that would make Canada's AI ambitions real instead of theoretical.
But when I evaluated them like I'd evaluate any provider we're trusting with production workloads, it fell apart.
No public pricing page. No SLA. No uptime guarantees. "Best effort" service — their words. No backups — also their words. Portal-based VM access. Website built on Wix.
I'm not saying this to be harsh. I genuinely hope they grow into something real. But we're planning training runs that cost five figures and take weeks.
"Best effort" and "no backups" aren't things I can bet my company on.
We emailed them asking if they can support production workloads at scale. I'm still hoping the answer is yes.
DigitalOcean: The Part I Still Can't Believe
Here's where the story takes a strange turn.
DigitalOcean — the company most people associate with $5 droplets and student projects — has H100 GPUs in their Toronto data center.
At $2.99 USD per hour.
They also have L40S at $1.57. RTX 6000 Ada at $1.57. RTX 4000 Ada at $0.76 — less than a dollar an hour for a GPU you can prototype on.
These were the Toronto offerings I recorded; listed offerings and prices did not guarantee capacity for a new launch.
A company best known for beginner-friendly web hosting has better Canadian GPU infrastructure than Google Cloud.
Google Cloud. A trillion-dollar company. With two Canadian regions. My procurement path was still unclear, despite the newer Toronto hardware described above.
Meanwhile, DigitalOcean — headquartered in New York, a fraction of Google's size — put H100s in Toronto. The catch is capacity.
There are reports of their Toronto GPUs selling out. Smaller providers have less inventory, and when it's gone, it's gone.
But at least the hardware exists. At least someone saw the demand.
UPDATE: April 16
Digital Ocean's GPUs are sold out - not just in Toronto, but in San Francisco, in New York, in anywhere.

However, I was asking sales about a 12-month plan and a reported $2.50-per-hour price; until accepted written terms and capacity were confirmed, it was a pending option.
I had contacted their sales team. That inquiry did not guarantee accepted terms or reserved capacity; the linked follow-up is the later account.
The Full Picture
Here's what GPU cloud shopping in Canada actually looks like as of the date of this writing:
Google Cloud Montreal:
April 6 check: P4/T4. April 14 update: L4 G2 added in zones b/c. The original search did not establish current launch capacity.
Google Cloud Toronto:
L4 and a listed A3 Edge H100 configuration. A3 Edge is inference-optimized; training suitability and launch capacity require a separate check.
AWS Canada:
Partial GPU support. Latest-generation training hardware limited and unclear compared to US regions.
OVHcloud Canada:
V100/V100S only. H100s are in France, not Beauharnois.
ISAIC Alberta:
Unverified. Non-profit university sandbox. No SLA, no backups.
DigitalOcean Toronto:
H100 at $2.99/hr, L40S at $1.57/hr, RTX 4000 Ada at $0.76/hr. The April 16 update reported on-demand availability unavailable in my search; a sales inquiry for reserved capacity was pending.
That was the set of providers I checked, not an exhaustive national inventory.
Why I'm Not Just Frustrated — I'm Worried
This isn't a niche infrastructure complaint. This is a structural problem.
In my search, I had not secured the Canadian capacity I needed. The fallbacks I was considering were these; they were not the only choices available to every Canadian startup.
Send the workloads I could not place locally to a US region, accepting that those workloads and their data would run outside Canada and that launch capacity still needed checking.
Or evaluate other Canadian configurations against the training job’s memory, interconnect, performance, price and availability requirements.
Neither option is acceptable for a country that says it wants to compete in AI.
Canada's federal government has an AI Compute Access Fund. It's a good program. I believe in the intent behind it.
For my project, funding alone would not solve the procurement problem unless I could actually obtain suitable Canadian capacity.
The offerings I had found still left me with quota, capacity or contract questions. That is not evidence that compute infrastructure was absent from Canada.
I sit in meetings where people talk about Canadian AI sovereignty. Everyone nods. Everyone agrees it matters.
Then I returned to my April Montreal search and still had not obtained the configuration I wanted. The April updates above record newer listings and explain why listing, account approval and a successful launch are separate checks.
The gap between the policy ambition and my procurement experience worried me. This diary did not measure whether that gap was closing nationally.
What I Think Is Happening
I don't think Google and AWS are ignoring Canada on purpose. I understand the business logic.
GPU hardware is scarce and expensive. Deploying thousands of H100s to a data center is a multi-million dollar commitment.
The US market is bigger. US demand is insatiable. If you're Google and you have 10,000 H100s to allocate, are you putting them in Iowa where demand is guaranteed, or Montreal where the market is smaller?
From their perspective, the answer is obvious.
I worried that relying on large providers alone would not give my project the access it needed; this search did not establish that competitive Canadian infrastructure could never emerge through market investment.
My view was that Canadian access should not depend entirely on decisions by Google and Amazon. That was a policy opinion, not a measured forecast of every provider’s future allocation.
Something has to change. Whether that's government incentives for domestic GPU deployment, Canadian cloud providers stepping up, or entirely new models for compute access — something.
For the training run and account I was evaluating, I had not yet secured suitable launch capacity. The newer hardware discussed above means this cannot be read as a national current-generation GPU inventory.
Where We Landed
We're building multi-cloud. DigitalOcean Toronto is our primary for Canadian workloads — if capacity holds. GCS and AWS in US regions as backup for training jobs that need scale we can't get domestically.
It's not what I wanted.
I wanted to build AI in Canada, on Canadian infrastructure, with Canadian data residency, supporting the Canadian AI ecosystem.
Instead, I'm relying on a New York company's Toronto data center as my best option and routing my largest jobs through Iowa.
I chose this country. I believe in it. I think Canada has real advantages in AI — the talent, the research community, the policy thinking.
But the hardware has to show up. The infrastructure has to be real, not marketing. The compute has to actually exist in the data centers, not just in the pricing pages.
That was the compromise I faced in my account and search. This diary does not establish the same experience for every Canadian AI founder.
The original checks were April 6, 2026, with April 14 and April 16 updates retained above. I consulted provider documentation and tested what my account could access; this post is not a complete launch-capacity measurement. Verify current terms, quotas and an actual launch before committing to a workload.
Read the update here for what I discovered three weeks later.
