TORONTO — While AI compute bottlenecks are getting better, some new startups still face challenges accessing processing power to develop their products. Investor Radical Ventures is trying to fill the gap, but it’s not simply stockpiling chips like some Silicon Valley peers. 

Instead, the Toronto-headquartered financier has arranged for founders enrolled in its masterclass program to get up to US$250,000 in credits from Google Cloud to spend on compute capacity and other tools, as well as support from the tech giant’s staff to set up their technology infrastructure. 

Radical is targeting academic researchers who want to turn their ideas into commercial applications, but haven’t yet set up companies, according to partner David Katz. “What we’re trying to do is address a gap in the market.” 

Researchers must currently compete for space on their universities’ server clusters, or muster the cash to buy their own chips. “For AI projects, particularly training runs, it’s not incredibly affordable,” Katz said. And while the cloud arms of Amazon, Google and Microsoft all offer free credits to startups, founders must first incorporate their firms to access them, Katz added. 

Radical says it doesn’t require Masterclass participants to take its money when their startups eventually fundraise, and the VC firm isn’t paying Google for the capacity. The investor hasn’t yet named the firms that have received the cloud credits.

Not all AI startups are compute-hungry, however. Few firms are building large language models (LLMs), which require huge quantities of processing power to train. Many of those already have tie-ups with cloud giants. 

Startups also need a lot of compute capacity if they’re developing smaller, specialized models for particular industries that have high performance requirements,  according to Kory Jeffrey, vice-president of technology at Inovia Capital, a Montreal-headquartered VC firm. Their founders tend to be machine-learning experts coming out of research labs. “You get a lot of PhDs that go that route,” Jeffrey said, although “as a percentage of the overall startup world, that’s actually quite small.”

Most new firms can instead build applications relatively cheaply using LLMs from providers like OpenAI or Cohere. Some AI executives also say compute is cheaper and easier to buy today than a year or two ago. Jeffrey said he has fielded calls from large and small cloud providers looking to rent hundreds of top-of-the-line Nvidia graphics processing units (GPUs) to Inovia portfolio companies. AI firms now “get bombarded with deals,” he said.

Some of the world’s largest tech investors now offer compute to portfolio firms. Silicon Valley giant Andreessen Horowitz reportedly plans to amass more than 20,000 GPUs, while Conviction and Index Ventures last year rented servers from cloud providers.

Canadian VCs haven’t gone that route. “Providers are generally eager to work with high-growth companies, and we’ve seen our portfolio companies navigate compute access effectively without needing much investor intervention,” said Allen Lau, operating partner at Two Small Fish Ventures, although he added the VC firm has the connections to help if necessary.

But Radical has been able to help the startups it backs by being proactive about cloud access, Katz said. “Really early, we saw what was happening in the compute landscape and how difficult it was to get access to state-of-the-art chips [and] pricing.” 

Radical has negotiated better terms with cloud providers for its portfolio firms, Katz claimed, including lower prices and shorter rental minimum periods. Its startups also get the flexibility to shift onto the latest chips that become available and to scale their usage up or down as needed. “Our companies are prioritized as target customers, and so they get great treatment,” he said.