Inside the fastest-growing Canadian AI startup you’ve never heard of

By Joe Castaldo · The Globe and Mail · July 3, 2026

The story of how Simon Eskildsen moved from Denmark to Ottawa and came to run one of the most intriguing companies in artificial intelligence today can be traced to the moment he dropped his iPhone as a teenager and broke the screen. Rather than buy a new one, he picked up an old Nokia phone and enjoyed a world free from a blinking, distracting device. “I feel it has had no major impact on my life to leave it behind,” he wrote on his website.

This was 2013, before tech detox memoirs were an established genre. His post surfaced on Hacker News, earned a shout-out from the New York Times, and, in his telling, caught the eye of Shopify recruiters who later offered him an internship. He had a precocious online presence, with blog posts about coding competitions and LinkedIn-style lessons from the 30 most productive days of his life. (“I quickly found out that I am already very productive and it proved difficult to cram in more things,” he wrote.)

He knew of Shopify but had never heard of its hometown of Ottawa. Still, at 18, he moved across the ocean to work as an intern, intending to complete a gap year. When he visited the office, however, he knew: “These are my people,” he recalled recently.

Mr. Eskildsen, 31, is still in Ottawa and assembling his own crew of like-minded people to build Turbopuffer Inc., one of the fastest-growing startups in Canada that you’ve probably never heard of, unless you’ve needed its services. The company has devised a new, efficient way for AI systems to search for information when serving up answers, a crucial feature for AI to be useful, while slashing costs for an industry that can’t stop losing money. It’s won Turbopuffer some massive customers, including Anthropic, the company behind Claude, now worth nearly US$1-trillion.

Turbopuffer is an unusual startup.

Mr. Eskildsen, its chief executive, started the company in 2023 with Justine Li, whom he met at Shopify. The pair could not appear more different.

Mr. Eskildsen has the ruddy mien of a Scandinavian athlete and appears to be vibrating with so many thoughts that his head is probably a few degrees warmer than it should be. Ms. Li, 31, has the quiet, wispy vibe of the kid trying to go unnoticed at the back of a classroom. What they share is a polymath’s ability to solve hard problems. “You could throw them into any industry, and they would succeed,” said Dale Neufeld, a former Shopify vice-president.

At less than three years old, their startup is on track to bring in more than US$100-million in revenue this year from about 1,200 customers, and it’s already profitable. While other startups need tens or hundreds of millions of dollars to scale up, Turbopuffer has required less than US$1-million in outside capital. The company has done this with 37 employees, most of whom are abroad, and a cramped Ottawa office down the hall from an online tire retailer, decorated with a solitary Lego-like pufferfish. “I would love to have the luxury of time to design an office,” Mr. Eskildsen said, “but it’s not worth the resources.”

Turbopuffer might sit more naturally in San Francisco rather than a few blocks from Parliament. While a lot of Canadian talent decamps for Silicon Valley, Mr. Eskildsen has no interest in leaving. He once spent six months in Berlin (“objectively one of the coolest cities on Earth,” he said) and missed running along the Rideau Canal, visiting Gatineau Park and hanging out with friends in the capital. “Turbopuffer has a very strong Canadian heart,” he said. “We want to invest as much as possible in Canada.”

Now if only he could sell everyone else on Ottawa.

If you ask Mr. Eskildsen what Turbopuffer does, you might get a lengthy lesson on the history of databases that spans hard disks, analytics, the challenges of compressing human knowledge and, eventually, AI.

The upshot is that because of how generative AI works, a new kind of database is required.

To start with, AI models do not understand language like we do. For them, text is chopped up into smaller bits called tokens and converted into long number sequences, called vectors, one of which might look like this: [0.91, -0.21, 0.65, 0.49…]. Now consider these number strings as co-ordinates plotted into a massive, multidimensional space. Words, phrases and concepts that have an association with one another (“dog” and “cat”) will be clustered nearby, while those that more rarely appear together in language (“dog” and “Weird Al Yankovic”) will be farther apart. Vectors capture context and meaning, so that a chatbot doesn’t confuse Dire Straits with the Strait of Hormuz.

Now let’s say you give a whack of PDFs to a chatbot so you can ask questions about them. The PDFs might be broken into chunks, converted to vectors and stored. Your queries become vectors, too. The system searches for vectors that are close to those from your questions, and serves up those portions of the PDFs as context for the chatbot to provide answers.

That’s one small example of a process called vector search. This feature explains how AI chatbots and agents can search through your conversation history, scour the web or pull internal documents and messages to surface the right information to inform answers. And it’s what Turbopuffer does for its customers. It combines this with more traditional text search features to catch elements that vector search might miss. Turbopuffer’s system isn’t limited to words, but any form of data that can be translated into vectors, such as images and audio.

The problem is converting information into vectors causes data to explode in size. “Dog” is not three letters: it’s a lot of numbers. A word or a phrase could run for thousands of numerical values. All of that data has to be stored somewhere and accessed quickly. The traditional way is expensive. Turbopuffer’s method is not.

To understand how the company did that, we have to talk about Mr. Eskildsen. He grew up the oldest of three kids in Aarhus, a coastal Danish city that he described as the “little brother” to Copenhagen. His father worked as a programmer before transitioning to management roles and his mother as a career adviser, while he was drawn to computers at a young age. He taught himself some coding and picked up more English through World of Warcraft than he did at school. A fellow player had a job designing websites for local businesses, and farmed out work to him. “I started programming and programming and programming,” he said. He took on more software projects from a local startup while juggling high school before adding another pursuit: studying for the International Olympiad in Informatics.

The elite competition challenges brilliant young people to code solutions to complicated problems that require careful study to even comprehend what is being asked. He powered his way through regional rounds to earn a spot to compete for Denmark internationally in 2012, and again the following year.

Here, too, he found his people. His mind was blown meeting kids younger than him who were, well, better at this sort of thing. The competitions spanned 10 hours over two contest days, with some participants sprinting to the bathroom so they wouldn’t waste problem-solving time. His own performance, he said, was “not good.” Still, the competitions equipped him to move to Ottawa when Shopify approached him in 2013. “I would not have moved to Canada at 18 if I had not had that experience,” he said. “I learned enough computer science to start understanding how all of it fit together.”

It was not unusual for Shopify to hunt for interns fresh out of high school, but it was unusual to recruit a teen from Denmark, a country better known for its shipping and pharmaceutical industries than for its software. “Our strategy with software developers was looking for enthusiasts, people who do it because they love it,” said Doug Tetzner, a former Shopify recruiter who interviewed Mr. Eskildsen at the time. “He fit the mould.” Once he was in Ottawa, it did not take long for Mr. Eskildsen’s gap year to become permanent. “It just felt like a place where I could fit in,” he said.

That meant skipping university, but he’s a voracious reader who regularly dumped information into a flash-card app. “Whenever there was something he didn’t know, he put it into this flash-card system, whether it’s what kind of tree that is, or which team won the Stanley Cup in some year,” said Ms. Li, Turbopuffer’s co-founder. Their shared interest in learning was one of the things they bonded over at Shopify, in fact.

Ms. Li grew up in Ottawa and loved tinkering with things as a kid, once cobbling together a night vision device to walk around her basement in the dark. She owns equipment for metal machining, welding and woodworking today. As a teenager, she toured the Shopify office with her mom and bumped into Mr. Tetzner, who offered her an internship the summer before she started studying computer engineering at the University of Waterloo in 2012. “She would sit and code all day,” Mr. Tetzner said, “just working uninterrupted for hours.”

She joined full-time after graduating and worked on user-facing features, but inched her way to Mr. Eskildsen’s projects because that seemed more interesting, she said. He worked on infrastructure, helping grow and maintain the complex plumbing powering Shopify as hugely popular flash sales strained its servers. “Simon and Justine spent years deep in the hardest parts of Shopify’s infrastructure, the stuff most people never see but millions of merchants depend on,” said Shopify president Harley Finkelstein in an email.

Keeping these systems functioning was so crucial that Mr. Eskildsen entered the orbit of founder Tobi Lutke. “That he was inner circle with Tobi is just a very good signal of the level that Simon operates at,” Mr. Tetzner said. One of Mr. Lutke’s beliefs as CEO, according to a 2014 Globe profile, is that people start with a trust battery at 50 per cent, and every decision they make fills or drains the battery. It was in this environment that Mr. Eskildsen felt comfortable enough to take power naps on a bench in full view of anyone walking by.

The confidence was earned. In the early 2010s, Netflix launched an internal program that randomly switched off servers to compel developers to build in resiliency. Mr. Lutke wanted to do the same thing. “It was a terrible approach for Shopify,” Mr. Neufeld said. Mr. Eskildsen came up with a different method, adopted by Shopify, that allowed developers to simulate failures in testing environments, more in line with company culture. “It just clearly demonstrates how principled a thinker he is,” Mr. Neufeld said.

After nearly eight years at Shopify, Mr. Eskildsen was ready to try something else. He left to start consulting, and took on work in 2022 from a startup called Readwise that made an app for storing content to peruse later. He’d been an early user, and became friends with the founders after emailing out of the blue years earlier. Now as a consultant, he looked into developing features such as article recommendations and a way for users to search their libraries, which is how, in a time before ChatGPT, he learned about vector search databases.

He also learned that deploying a vector database would be obscenely expensive, costing Readwise tens of thousands of dollars – each month. As a bootstrapped startup, Readwise shelved the plans. But Mr. Eskildsen could not let it go. “I think he really took it personally,” said Readwise co-founder Tristan Homsi. “Like, ‘How am I not able to build this?’”

Mr. Eskildsen was determined to find a less expensive method. He will approach engineering issues with what he calls napkin math, wherein he games out the expected performance of a system with quick calculations. (He has a series of blog posts on the topic.) He applied the concept to vector search. In a way, it was like a challenge from a coding competition: solve for X number of vectors at the lowest cost. “The architecture of all the other databases was the same,” he said. “It was like, how are they going to compete with each other?”

Database providers relied heavily on random access memory, or RAM, to store and process vector data. It’s fast, but expensive. Mr. Eskildsen instead focused on something called object storage, which is ideal for hosting lots of data at a fraction of the cost. Think of it like this: If you’re cooking, you might lay out some ingredients on the countertop within arm’s reach, but space is at a premium. That’s RAM. Everything else is in the pantry, which has more room, but takes longer to fetch. That’s object storage. (Though the difference in speed between RAM and object storage is so vast that in this analogy, your pantry would have to be somewhere in the vicinity of Mars. And it’s not a pantry but an Amazon warehouse.) After a few months, Mr. Eskildsen developed an architecture that keeps vector data in cheap object storage and selectively loads it into a cache (a faster storage layer) and to RAM when needed. “If you’ve been building databases for 30 years,” he said, “it’s very difficult to retrofit this.” A completely new query through Turbopuffer is slower than if everything were kept in RAM, but he wagered that would be an acceptable trade-off given the cost benefits. (Plus, once data is in cache or RAM, the system is as fast or faster than other databases.)

In October, 2023, he wrote on X that he had an option that was up to 70 times cheaper than existing vector databases. “If you’re interested,” he wrote, “DM me.”

AI coding company Cursor is worth about US$60-billion today, or at least that is what SpaceX agreed to pay in June to buy the San Francisco-based company. But when it got in touch with Mr. Eskildsen after his post, it was a little-known startup buckling under the weight of its own growth.

“Their infrastructure was breaking,” he said. He knew he could drastically lower Cursor’s vector search costs – and that he would need help. So in October, he asked Ms. Li, who had left Shopify, to join him as a co-founder. “Justine would take it to levels that I just couldn’t. She’s a much better programmer,” he said. Turbopuffer was incorporated that month.

With Ms. Li on board, he flew to San Francisco in November to meet Cursor’s four co-founders, all men in their 20s, at least two of whom read his napkin math posts. He stayed late at their office, having helped with an unrelated database issue, and realized nobody seemed to be going home. Still on Eastern time, he craved sleep. Showing weakness in front of hard-core coders was unwise, he felt, so he feigned an excuse about meeting a friend and slipped out. He had already committed that day to slashing Cursor’s vector database bill by 95 per cent – which is what happened when Cursor migrated to Turbopuffer a week or so later, becoming its first customer.

Around that time, Mr. Eskildsen was introduced to Lachy Groom, an angel investor in San Francisco who by his own admission does not have much of an investing thesis. “I meet people and I try to form an opinion of whether this could be someone that builds an extraordinarily important and valuable company,” he said. They met in the California city in November, and on a long walk, he was struck by Mr. Eskildsen’s intelligence, but also that he didn’t seem to want money. “The idea of him taking money was just very uncomfortable. I felt almost uncomfortable pushing the idea of it,” he said. Mr. Eskildsen wanted just enough to get traction, and offered to return some or all of the funds if it didn’t work out. “You never see that in Silicon Valley,” Mr. Groom said.

As word travelled about Turbopuffer, venture capitalists started salivating, but Mr. Eskildsen proved elusive. Avery Klemmer at Thrive Capital in New York messaged through X and got no response, but eventually snagged him for a Zoom call through connections in her network and flew to Ottawa in May, 2024. “They like to keep their head down and not be a part of the hype cycle,” she said. At the time, vector database companies such as Weaviate and Pinecone were already up and running. “Simon held the floor and explained his thesis for why Turbopuffer should exist,” she said. “We meet founders all day. There’s just not that many people who are highly articulate about why they’re building.”

Thrive invested in December, 2025, and that might be the last outside funding Turbopuffer needs. All told, it has raised between US$700,000 and US$900,000. Mr. Groom doesn’t think more will be necessary since the company is profitable and growing fast, though that hasn’t stopped investors from pinging him in hopes he can make an intro to Mr. Eskildsen. “People literally fly to Canada and message him,” he said.

Turbopuffer has not had to sell itself too hard to find customers, either. It counts AI companies such Harvey and Legora on its roster. (Readwise became a customer, too.) For any outfit with large search workloads, the price is compelling. That was the case for Telus Corp. in 2023, when it was building a platform for companies to access a variety of AI models. “We were pretty close to giving up,” said Justin Watts, an engineer at Telus. A vector database was looking cost-prohibitive at $200,000 per month; Turbopuffer could do it for $5,000.

Mr. Watts felt like he almost had to convince Mr. Eskildsen to take the company as a customer, though. “He was like, ‘Are you going to be worth my time?’” he said. Mr. Eskildsen, who projects a self-assured cheekiness, wanted huge workloads to push the limits of his system. As for Telus, the biggest risk was trusting a startup no one had heard of. “I put a lot of my credibility on the line,” Mr. Watts said. It became clear there was no reason to worry after Telus became a customer in early 2024. “It just worked. And then the bill was a joke. We felt like we were robbing Simon,” he said.

Turbopuffer’s highest-profile customer is Anthropic. Mr. Eskildsen recognized in the AI giant could be key when OpenAI purchased a company in 2024 to improve its retrieval capabilities. Mr. Groom made the introduction, and Anthropic signed as a customer last year. No one will say what exactly Turbopuffer is doing for Anthropic, which did not reply to requests for comment, though Ms. Li gave some hints. “They are doing really large-scale search on our system,” she said. “We’re actually powering one of the key pieces of infrastructure at Anthropic.”

While Turbopuffer’s flagship customers are essentially borne of AI, that is changing as adoption grows. Any outfit with lots of data could benefit from connecting to AI systems for search. “If you pick any Fortune 100 company, there’s reams of data sitting there, ready to be ingested,” said Nikhil Benesch, Turbopuffer’s chief technology officer.

The thing about technology is that it moves quickly. Turbopuffer is proof of that. But it cuts the other way, too. The company’s method is not impossible to replicate or improve. “It’s a well-executed recipe of known ideas,” said Andy Pavlo, associate professor of computer science at Carnegie Mellon University. Amazon Web Services, for one, debuted a service that sounds remarkably like Turbopuffer last December.

Turbopuffer had a soul-searching crisis a few months into its existence, when an established competitor launched a similar offering. “We got really scared. We were like, should we pivot?” Ms. Li said. Instead, Turbopuffer pushed aside some feature ideas to double down on its core offering, and hasn’t looked back. But it’s not an experience she will forget. “A competitor can come out of nowhere at any time and launch something that is going to blow you out of the water,” she said.

To keep swimming, Turbopuffer has to ensure its system performance never slips. “Simon’s hard-core engineering background can tweak it and make sure that it runs really, really fast,” Prof. Pavlo said. Mr. Eskildsen does have an obsessive bent. In a conversation with one of Cursor’s co-founders earlier this year, he recalled how a three-second load time for a Shopify store in Australia once “haunted” him. His napkin math, you see, indicated it should be in the order of milliseconds.

Companies – especially startups – are often reflections of their founders, and that is true of Turbopuffer. Mr. Eskildsen described the culture as “hard core and whimsical,” the kind of place that hires exceptional people, but is infused with a hint of playfulness.

The initial spark for putting “puffer” into the company’s name, for example, occurred when he and a friend were riffing and took inspiration from a pufferfish emoji. He once nervously joked to Nvidia founder Jensen Huang at an event that he named the company turbopuffer – branded nonchalantly in lower-case – so it could pivot to vaping if things didn’t work out.

Canada is an important part of Turbopuffer’s DNA, too, even though most of its employees live elsewhere. Earlier this year, the company held a retreat in Banff, Alta., and Mr. Eskildsen sounded genuinely chuffed that a couple of employees from California and India got to experience snow for the first time.

There is a lot of concern in Canada about losing founders to the U.S., particularly in AI, evidenced by the $500-million fund announced by the federal government in June to invest in promising companies. It could be a lot easier for Mr. Eskildsen to leave for San Francisco – “That’s where the customers are, and I’m disadvantaged by not spending more time with them,” he said – but he’s still in Ottawa. Married with one child, he’s built a life there.

Turbopuffer is the kind of company Canada needs to fulfill its AI ambitions, so it’s fair to ask what led to its creation and whether it can be replicated. The reality is it has little to do with government policies to attract and support talent. In fact, Turbopuffer applied to the federal government’s flagship AI fund to subsidize compute costs only to get rejected without explanation this year.

Instead, Turbopuffer’s reason for being here has more to do with chance – Mr. Eskildsen broke his phone, Shopify reached out, and he found Ottawa to be a good place to live. (The latter reason should not be discounted when crafting policy; working to improve quality of life benefits everyone, not just entrepreneurs.) The reason Shopify is in Ottawa is just as random: Founder Tobi Lutke, who is German, met a woman from Ottawa on a trip to British Columbia, fell in love, got married and, for a time, lived in her family’s home in the capital.

Some of Canada’s biggest tech successes might just be happy accidents.

There are ways to increase the odds, and naturally Mr. Eskildsen has ideas. He has a list titled “What should Canada do?” that includes tax changes, daily flights to U.S. tech hubs and a program to expedite immigration visas for qualified companies. Closest to his heart is for Canada to invest more in math and science competitions, like his beloved International Olympiad in Informatics. “In my perfect world, these Olympiads are treated with the same kind of status around the world as the sport Olympics,” he said. “These people go on to do pretty remarkable things.”

He noted that Turbopuffer’s first hire was Bojan Serafimov, an IOI silver medal winner from North Macedonia. But such people – his people – are rare. While he will hire the best talent regardless of where they live, he’s found it “extremely difficult” the past few years to find people willing to move. Of the company’s 37 employees, 11 are in Canada, with others in the U.S. and Europe. This could be one problem beyond the scope of napkin math.

Mr. Eskildsen is trying, though. He’s offering incentives to employees if they relocate to Ottawa. No one has taken him up on it, but he’s hopeful. “There’s murmurs,” he said. “I will keep selling.”

With reports from Sean Silcoff