Build North
Subscribe
← All articles
AIOct 25, 20268 min read

Neoclouds are AI's landlords. Canada's problem is megawatts

By Jatin Rajvanshi, product manager in Vancouver

A new kind of cloud company now rents out much of the world's AI computing power. Here's what neoclouds are, why their finances make investors nervous, and what Canada's power grid has to do with your AI roadmap.

If you've used an AI product this year, there's a good chance the model ran on a computer owned by a company you've never heard of. Not Amazon, Microsoft or Google, but a "neocloud": a cloud company that does one thing, rent out GPUs for AI.

The biggest, CoreWeave, brought in US$2.58 billion in revenue last quarter, more than double a year earlier, and has a backlog of about US$104 billion in signed contracts. Lambda's backlog jumped from US$15 billion to US$50 billion between June and September, mostly thanks to a single US$35-billion commitment from Anthropic.

This month, the mood turned. CoreWeave's bond yields climbed into the 15% range, and investors started asking whether the whole model is built on borrowed money.

Here's our take for Canadian builders: neoclouds are a real and useful shift, and they're often cheaper than the big three. But don't treat them as the safe default. And in Canada, the thing that will decide who gets AI computing power isn't chips. It's electricity.

The TL;DR:

  • Neoclouds (CoreWeave, Lambda, Nebius, Crusoe and others) are GPU-first clouds. List prices for top AI chips can be 40% or more below Amazon's.
  • Their growth runs on debt and a handful of giant customers. That's a risk you inherit if you build on one without an exit plan.
  • Canada has its own players (Bell AI Fabric with BUZZ HPC, Telus's AI factories), but American hardware and owners are everywhere in the stack.
  • The real Canadian bottleneck is grid power. Alberta has requests for far more power than its grid can supply, and Quebec has paused large new data centre power deals.
  • If you're an AI startup, Ottawa's AI Compute Access Fund covered two-thirds of eligible costs on Canadian clouds. Watch for the next intake.

What a neocloud actually is

Amazon Web Services, Microsoft Azure and Google Cloud (the "hyperscalers") sell you hundreds of services: databases, storage, email, AI. A neocloud mostly sells one: access to large clusters of Nvidia GPUs, set up for training and running AI models.

That focus shows in the price. In an October comparison of public list prices, Lambda charged US$3.99 per hour for an Nvidia H100 GPU in an eight-GPU machine, versus US$6.88 on Amazon and about US$12.29 on Azure's eight-GPU option. Regions and setups differ, so it's not a perfect comparison. But the gap is real.

Don't assume prices only go down, though. The same comparison found Nebius raised its H100 rate from US$3.85 to US$4.50 on October 1, with increases at several other providers too. When demand is this strong, compute behaves like a scarce commodity.

The business model makes investors nervous, and it should make you think

A neocloud's business is simple to describe: borrow money, buy GPUs, rent them out, and hope the rent covers the loan before the chips go out of date.

The numbers show the strain. CoreWeave's revenue doubled last quarter, but it still lost US$626 million, with US$640 million in net interest costs. Futuriom reports that CoreWeave plans to spend more than double its expected 2026 revenue on new infrastructure, and that Nebius's spending could exceed three times its annual run-rate revenue.

Three things are worth understanding:

  • Customer concentration. Lambda's backlog story is mostly one customer, Anthropic. If a handful of AI labs slow down, the whole sector feels it.
  • Circular money. Nvidia has invested in CoreWeave, Crusoe, IREN, Lambda, Nebius and Nscale, and has agreed to buy CoreWeave's unused capacity through April 2032 in a deal initially worth US$6.3 billion. The chip maker is propping up its own customers.
  • Chip ageing. Many providers spread GPU costs over six years. Futuriom argues GPUs may really lose their value in four to five, which would make today's profits look better than they are.

None of this means neoclouds will collapse. CoreWeave's adjusted EBITDA was US$1.5 billion last quarter, so the core rental business makes money before interest and depreciation. But if you're a startup signing a multi-year GPU contract, you're making a bet on your provider's balance sheet. Price it in.

Canada has its own version, built on borrowed parts

Canada is building sovereign AI computing that looks a lot like a neocloud, often run by telecom companies.

In June, Bell announced that Cohere will supply AI tools through BUZZ HPC's cloud, running on Hypertec hardware in an expanded Bell data centre in Merritt, B.C., with Nvidia chips. Bell's CEO said, "Too much of our AI computing capacity is being served from other places, particularly the U.S." Telus runs a "sovereign AI factory" in Rimouski, Que., and in May Ottawa and Telus announced talks on a large sovereign AI data centre in B.C., though no federal funding has been committed yet.

Look closer, though, and the American parts are everywhere. In 2024, Ottawa gave Cohere $240 million to help pay for computing power from a CoreWeave data centre in Cambridge, Ont., and CoreWeave is also a tenant at Bell's planned 300-megawatt site in Sherwood, Sask. BetaKit reported that Telus uses Nvidia hardware, Bell chose Groq for its B.C. data centres, and American firms own nearly a third of Canada's 283 data centres. The same piece notes that ThinkOn was the only Canadian company on the federal government's shortlist of approved cloud providers.

That matters because "the data is stored in Canada" isn't the same as "the data is under Canadian control." Experts quoted by BetaKit point to the U.S. CLOUD Act, which lets U.S. authorities request data from U.S.-owned providers wherever it sits. If your customers are in health, finance or government, they'll ask who owns the company running the GPUs, not just where the building is.

Canada's real bottleneck is power, not chips

You can order GPUs. You can't order a grid connection.

Canada had about 1.4 gigawatts of live data centre capacity as of September 2025, compared with 40.6 GW in the U.S., according to law firm BLG. For scale, CoreWeave alone reported about 1.5 GW of active power in June.

The provinces are rationing:

Canada still has real advantages: a cold climate, lots of hydro power and proximity to the U.S. market. But for the next few years, the provinces that can connect new demand fastest will decide where Canadian AI computing gets built.

What you can do tomorrow

Founders

  • Before signing a long GPU contract, ask who the provider's biggest customers are and how it's financed. Keep your workloads portable (containers, open tooling) so you can move if prices or providers change.
  • Watch for the next intake of Ottawa's AI Compute Access Fund. It's closed now, but it covered two-thirds of eligible compute costs on Canadian cloud services and half on foreign ones, for projects of $100,000 to $5 million. That difference can make a Canadian provider the cheaper option.

PMs

  • Add "where does this run, and who owns it?" to your AI feature reviews. Your enterprise and public-sector customers will ask.
  • Budget for compute prices that can go up. Build in room for price changes when you set AI feature pricing.

Engineers

  • Benchmark one neocloud against your current hyperscaler on a real workload. Measure cost per useful result, not just the hourly rate, and include data transfer costs.
  • Keep a second provider tested and ready. A capacity crunch is a bad time to learn a new platform.

Where I think this goes

The whole neocloud model is a bet on one thing: demand for GPUs and compute. Right now, I think most companies are still in a learning phase. They're working out what AI can actually do for them, and honestly asking why they'd need this much compute at all. That uncertainty is what you're seeing in the debt markets this month.

So I wouldn't be surprised by a slump in the middle: a stretch where capacity runs ahead of demand, prices soften and some of the most indebted players struggle. But I think it's temporary. Once companies figure out what to do with this much compute, over the next year or two, I expect demand for neoclouds to go up, not down. If that dip comes, Canadian builders should use it to experiment while compute is cheaper, not read it as the end of the story.