VESSL AI

Here is a boring fact that turns out to be a business: most of the GPUs a company pays for are, at any given moment, doing nothing. They sit in a rack, or in someone else's cloud, warm and idle and expensive, while the finance team wonders why the AI bill looks like a mortgage. VESSL AI was built on that gap between what you pay for and what you use, and its entire pitch can be compressed into a single unglamorous sentence: we will find you a cheaper chip and send your job there automatically.

That sentence, of course, is doing a lot of work. Underneath it sits an MLOps platform - the machinery that lets machine-learning teams train, deploy, and monitor models without personally negotiating with Kubernetes every morning. VESSL pools GPU capacity across on-premise clusters and multiple cloud providers, then routes each workload to whichever resource is cheapest and available. The company describes the result as a "GPU liquidity layer," which is finance-brain language for treating compute like capital: something you move around to where it earns the most, rather than something you buy and let depreciate in a corner.

The headline claim is that this can cut GPU spend by up to 80%. That "up to" is carrying a suitcase, as all "up to" numbers do - the savings depend on how much you were overpaying to begin with, and the trick is mostly a combination of multi-cloud arbitrage and spot instances, the discounted, interruptible compute that cloud providers sell when they have spare capacity. Spot instances are cheap because they can be yanked away mid-job; the value VESSL adds is the orchestration that makes that acceptable, with auto-failover and traffic routing so a reclaimed instance is an inconvenience rather than a catastrophe. If you have ever watched a three-week training run die at hour 400, you understand exactly why someone would pay for this.

01Who built it

VESSL AI was founded in 2020 by four engineers - Jaeman Kuss An, Jihwan Jay Chun, Intae Ryoo, and Yongseon Sean Lee - whose résumés read like a tour of places where scale hurts. Among them: Google, the AI-startup world, and PUBG, the studio behind the battle-royale phenomenon. An, the CEO, spent roughly seven or eight years as a software engineer, at one point building a mobile game that reached ten million users, which is the kind of experience that teaches you, viscerally, what happens to infrastructure when demand arrives all at once. He later worked at a medical-AI startup building a deep-learning system to predict acute illness in hospital patients. The through-line is not any single domain but the recurring pain of making models run reliably on hardware that would rather not.

The team is unusual in its geography: it runs across two continents, with people in Seoul and in the San Francisco Bay Area, roughly sixteen time zones apart. That is either a coordination nightmare or a competitive advantage depending on the week, but it reflects a deliberate bet - Korean engineering roots, American market. The company has said it wants around half its revenue to come from the United States.

02What you can actually do with it

VESSL's product line is modular, and each piece is named with refreshing literalism. VESSL Run is the atomic unit - a single Kubernetes-backed workload that handles training and experimentation. VESSL Serve pushes models into production with real-time serving, traffic splitting, and auto-scaling. VESSL Pipelines strings together the messy multi-step workflows - LLM fine-tuning, data preprocessing, batch inference - that real ML work actually involves. VESSL Cluster manages GPU resources across the hybrid on-prem-and-cloud setup, and VESSL Hub offers one-click recipes for deploying open-source models like Llama, Mistral, and Stable Diffusion.

The newest piece, Hyperpocket, is a different sort of move. Launched in early 2025 and given away as open source, it lets developers connect their own tools to any AI agent framework - addressing the very practical limitation that a clever agent is useless if it cannot actually do anything in the world. Giving away the connector while selling the compute is a recognizable strategy: the free thing manufactures demand for the paid thing.

03Who's paying

The customer roster is the most interesting tell, because it is so incongruous. On one list you will find Hyundai, the mobility joint venture TMAP, Hanwha Life Insurance, and the Korean AI startups Upstage, ScatterLab, Yanolja, and Wrtn.ai. On the same list sits LIG Nex1, a defense and aerospace manufacturer, alongside academic users at Stanford, Carnegie Mellon, Columbia, Duke, Michigan, and Washington. A grad student training a model and a defense contractor shipping one want, it turns out, precisely the same thing: for the GPUs to just work. Roughly 50 enterprise customers and more than 2,000 users, at last count.

04The money

In October 2024, VESSL closed a $12 million Series A, bringing total funding to roughly $16.8 million. The round was led by a syndicate of largely Korean investors - A Ventures, Ubiquoss Investment, Mirae Asset Securities, Sirius Investment, SJ Investment Partners, Wooshin Venture Investment, and Shinhan Venture Investment - and earmarked for the infrastructure needed by companies building custom LLMs and vertical AI agents. It is not an enormous war chest by Bay Area standards, which makes the cost-efficiency pitch feel less like marketing and more like autobiography: a company that sells frugality has some incentive to practice it.

05Where it sits

VESSL occupies a crowded and fast-moving neighborhood. On one flank are the MLOps platforms - Weights & Biases, Determined AI, Anyscale, and the hyperscaler suites like AWS SageMaker and Google Vertex AI. On the other are the "neo-clouds," the GPU-rental specialists such as CoreWeave, Lambda Labs, and RunPod, plus inference shops like Together AI. VESSL's positioning is to be a bit of both at once: the workflow software and the compute routing, bundled, with the multi-cloud story as the differentiator. Whether that bundle wins is an open question, but the company has collected the kind of validation that matters in this world - a top-five finish in NVIDIA's Inception Startup Grand Challenge in 2025, and a CES 2025 showing of MLOps-powered AI agents that drew interest from global big-tech firms.

None of it is guaranteed. The GPU market is volatile, the incumbents are enormous, and "up to 80%" is a number that will eventually be tested by a skeptical procurement department. But there is something durable about building the unglamorous pipes before everyone realizes they are thirsty. VESSL started laying MLOps plumbing in 2020, back when "custom LLM" was not yet a phrase anyone said out loud. Four years later, it is exactly the plumbing the boom needs - and the company's whole reason for existing still fits in one sentence about not wasting money.