Profile Founders • GPU Infrastructure
He was two credits from finishing at Stanford. He left to build TensorPool, a tool that turns the headache of cloud GPUs into a single line you type and forget. The bet: make industry-scale compute feel like coding on your own laptop.
There is a specific kind of nerve it takes to walk away from a Stanford degree when you are two credits from having it in hand. Not two years. Not a full semester. Two credits - the academic equivalent of standing on the doorstep and deciding not to knock. Tycho Svoboda did exactly that, and he did it to run a company that fits, more or less, inside a single command you type into a terminal.
The company is TensorPool, and the pitch is almost rude in its simplicity: describe the machine learning job you want to run, hit run, and let the software find the cheapest cloud GPU on the market and do the rest. No servers to configure. No SSH sessions. No moving mountains of data around. As co-founder and CEO, Svoboda has a shorthand for it that borrows a name every developer already knows - "Vercel for GPUs."
Most good infrastructure companies start as somebody's personal annoyance, and TensorPool is no exception. Svoboda and his two co-founders trace the frustration all the way back to high school, when they trained their first models and ran straight into the wall that every ML practitioner knows. Renting remote compute was a chore. Amazon's EC2 wanted an intimidating amount of configuration. Google Colab wouldn't keep a job running if your laptop went to sleep. Years passed, they moved into academia and industry, and the wall never really moved.
"We wanted using GPUs to be as easy as programming locally."
Tycho Svoboda, co-founder & CEO, TensorPoolSo they built the thing they wanted. The product has two moving parts, and both are aimed at the same target - friction. The first is pure developer experience: you stay inside your IDE, TensorPool ships your code to the GPUs, runs it, and sends the results back as if it had all happened on your own machine. You can spin up several experiments with different config files and version-control your training runs the same way you version-control code.
The second part is the money part, and it is the cleverer of the two. Cloud GPUs come in two flavors. Spot instances are cheap but can vanish out from under your job at any moment. On-demand instances are dependable but expensive. TensorPool runs a real-time analysis across the available GPU cloud providers, places your job on the cheapest one, and uses what the team calls spot node resumption to give you spot-instance prices with something close to on-demand reliability. The headline result is roughly half the cost.
For a young company selling to machine learning teams burning cash on compute, that is a clean wedge. The cost line is the thing everyone feels, and cutting it in half is a message that needs no translation. It also fits the moment - GPUs have been scarce and expensive, and the market has been long on great chips and short on painless access to them.
TensorPool is not a solo act, and its origin story is almost suspiciously tidy. Svoboda met his co-founders, Joshua Martinez and Hlumelo Notshe, as freshmen at Stanford, where the three were placed on the same dorm floor. From there they took essentially all their classes together, and over the following years they scattered into the kind of resumes that make recruiters sit up - Apple, DeepMind, Blackstone, Nextdoor, the Chan Zuckerberg Initiative, PIMCO. Then they came back together to build one thing.
Svoboda's own path into infrastructure came sideways, through data. Before he was orchestrating GPU clusters he was working data analytics roles - a Blackstone internship, stints at Blumen Systems, PIMCO and MBF Bioscience, and a turn as a venture partner at Pioneer Fund. It is not the classic hardware-hacker origin, and that may be part of why the company leads with cost and ease of use rather than raw technical bravado. He learned early to look at where the money goes.
Ask Svoboda why he'd leave school this close to the finish line and the answer is less about TensorPool specifically and more about tempo. He describes the current AI landscape as "such a freakishly fast-paced world," the kind where a six-month head start feels like a year. His justification for going all in is disarmingly blunt about the season of life he's in.
"I don't have a wife, I don't have a kid, I can focus every single ounce of my life into this."
Tycho Svoboda, on why nowHe is also self-aware about how young this cohort has become. At Y Combinator events where alcohol is served, he noticed what he called a "scary amount" of founders wearing badges marking them as too young to drink - and cheerfully admitted he'd have been wearing one himself not long ago. "If you caught me six months ago, I would have had one of those stickers too," he said. By his count, at least a dozen of his Stanford classmates have also dropped out to build AI companies. The doorstep he stepped off of, in other words, is getting crowded.
That crowd is the backdrop that makes his choice legible. In a normal year, walking away from a nearly-finished Stanford degree reads as reckless. In this one, surrounded by peers making the same math, it reads as a rational response to a window that everyone in the room believes is open right now and might not stay that way. Whether the window is as narrow as they think is the open question - but conviction, not caution, is what builds infrastructure companies in their first year.
TensorPool launched publicly through YC's Winter 2025 batch with the founders doing what founders do - handing out free compute to early users and asking, plainly, for people to try it and for intros to enterprises that might need cheaper GPUs. The company's tone is friendly and terminal-native: get set up in three minutes, run one command, drop a star on the GitHub repo while you're there. There's a short demo video of the tool fine-tuning a model, which is the fastest way to understand why the pitch lands.
What Svoboda is really selling isn't a discount. It's the removal of a decade-old chore - the SSH sessions, the config sprawl, the data migrations that stood between an idea and a trained model. If TensorPool works the way the pitch promises, the win is measured in the thing every builder actually wants back: the hours between typing an idea and watching it run.