Run:AI

The problem: expensive cards sitting idle

Run:AI was founded in 2018 in Tel Aviv by Omri Geller and Ronen Dar, and its whole product grows out of a fairly simple organizational problem. A large organization that develops AI has several teams, and each team is assigned graphics cards: compute accelerators that cost thousands to tens of thousands of dollars each. They are allocated by team, not by actual load, so the result is familiar to anyone who has run a department. One team waits in a queue for days to train a model, while right next door in the server room another team's cards sit unused. At a time when advanced accelerators were not only expensive but actually in short supply, this waste turned from an operational problem into a first-order budget problem. The question the company asked was not how to build better hardware, but how to make the hardware that already exists run closer to its full capacity. It is worth showing the scale. Training a single model can keep dozens of accelerators busy for days on end, while a small research experiment needs one accelerator for an hour. When all of these share the same pool of hardware, the gap between equipment that works most of the time and equipment that sits idle most of the time translates directly into budget.

The solution: a management layer on top of the hardware

Run:AI built a software layer that sits between the users and the hardware and manages access to it. Instead of each team holding its own cards, all the cards go into a shared pool, and the software divides them dynamically according to what is running at any given moment. A few key mechanisms make this work. Fractional sharing lets several small jobs share one accelerator instead of each taking a whole one. Queues and quotas make sure every team gets its fair share over time, even if at a given moment someone else is using more. And gang scheduling makes sure a job that needs several accelerators at once gets all of them together or does not start at all, because such a job that gets only some of the accelerators will get stuck and hold them for nothing. All of this is managed on top of Kubernetes, the standard orchestration system for managing workloads in the cloud. For the user, the result is a change in how work feels, not just in the budget line: instead of requesting a fixed allocation in advance and waiting for it, the researcher sends a job to the queue and gets resources as they free up.

The funding rounds and what came before the acquisition

In March 2022 the company completed a $75 million Series C round, led by Tiger Global Management and Insight Partners with participation from existing investors, bringing its total funding to $118 million. In the round announcement, the company reported a ninefold increase in annual recurring revenue in the year before, and a workforce that more than tripled over the same period. These figures came from the company itself, as is usual in funding announcements. The customers it had gained by then were large organizations in finance, health care and autonomous vehicles, exactly the kind of organization that runs its own in-house computing infrastructure rather than only renting it in the cloud. This scale, a little over a hundred million dollars rather than billions, matters for understanding what happened next, because it places Run:AI as a focused software company rather than a large independent infrastructure player. A caveat is in order: figures like these, published in a funding announcement, are not audited and are always presented in the most flattering light. A ninefold increase in recurring revenue sounds especially impressive when the starting point is low, which is usually the case for a company of this age.

Nvidia's acquisition and the regulatory review that came with it

In April 2024 Nvidia, the world's leading maker of compute accelerators, announced its intention to buy Run:AI. The deal's value was not officially disclosed and was reported in the business press as about $700 million. The acquisition was actually completed only on 30 December 2024, about eight months after the announcement. The gap came from regulatory reviews: both the European Commission and the US Department of Justice opened separate reviews of whether the deal harmed competition, and the European Commission approved it in December 2024. It is important to understand why a deal of this size, small compared with major tech acquisitions, drew a review at all. Run:AI's software had until then worked only with Nvidia hardware, and regulators were concerned that a buyer that already held a dominant market share would also get the management layer above it. This made the case a well-known example of how the size of a deal is not what decides how hard it is reviewed. The eight months between the announcement and the closing illustrate a rule that comes up in many entries here: an announced acquisition is not an acquisition. Until all the approvals are given and the money has actually changed hands, it is a stated intention that can still change.

The code goes open: KAI Scheduler

On 1 April 2025 Nvidia made a move that was not a given: it released the scheduling engine that had been the core of Run:AI's product as open source, under the name KAI Scheduler and the Apache 2.0 license. The official explanation was to make it more widely available to the whole AI ecosystem. The open-source project includes the same mechanisms described above (fractional sharing of accelerators, queues and quotas, gang scheduling), and it continues to be packaged and sold as part of Nvidia's commercial software platform at the same time. Technically, the move also makes it possible to adapt the scheduler to other makers' hardware. It can be read in two ways, and probably both are true at once: a gesture to the community that increases adoption, and at the same time an answer to some of the competition concerns regulators raised in that same review. For developers around the world the meaning is very practical: a scheduling engine that used to be a paid commercial product became free to download and use, including in small organizations that would never have bought it anyway.

What this story teaches

Run:AI did not build any hardware. All of its value lay in the software layer on top of hardware someone else made, and yet it became a component in the platform of one of the most valuable companies in the world. This is a recurring pattern in this industry: when a compute resource becomes scarce and expensive, whoever knows how to share it efficiently creates real value, sometimes no less than whoever makes the resource itself. There is also a side worth stating plainly: the fact that the software worked from the start only with one maker's hardware was both its strength with that maker and its dependence on it. For the Israeli industry, the reported deal value, hundreds of millions of dollars for a company founded six years earlier, strengthened Tel Aviv's standing as a center for building AI infrastructure, alongside other companies in the same field. There is also a lesson here for small software companies. Being an essential component in a large player's infrastructure is a legitimate and rewarding path, but it almost always ends in an acquisition, because a company whose infrastructure rests on you will prefer to own you rather than rely on you.