Before there were model weights and GPU clusters, there was the inbox. In the early 2000s, it was a crowded and unruly place. A single person could mark a message as spam, but the next person would still have to make the same judgment. Vipul Ved Prakash saw a systems problem hiding inside a daily irritation. He wrote Vipul's Razor, an open source filter that let users identify unwanted mail together. The name was playful; the premise was serious. A network could be smarter than one isolated computer.
That small act of collective defense is a useful opening to Prakash's story. He now runs Together AI, a company that helps developers train, customize and run open AI models. The machinery is vastly larger than an email filter. Yet the old question persists: what happens when a powerful capability is shared rather than kept behind one door? Over a career spanning spam, social search, Apple and AI infrastructure, he has approached that question with the habits of an engineer and the patience of a founder.
A college exit, with a reason
Prakash grew up in New Delhi. He attended St. Stephen's College, studying mathematics, physics and computer science, but left before completing a degree. A 2003 account recorded his explanation in wonderfully specific terms: he wanted “undisturbed coding time.” It is a better detail than the familiar mythology around college dropouts. The point was the work itself, and the hours it demanded.
In the late 1990s he helped start Sense/Net, an early internet venture in India. He wrote code, including open source Perl extensions, and later studied authorship across millions of lines of Linux code. Those were years when the internet felt both newly public and newly messy. Spam was an everyday proof that a network could be useful to everyone and exploitable by anyone. Razor answered with cooperation: if enough users reported the same unwanted message, the system could recognize it for others.
In 2003, MIT Technology Review named him among its young innovators for the anti-spam work. Recognition arrived early, but the more revealing development was commercial. He co-founded Cloudmark, which developed tools to protect messaging systems. The idea that started in open code now needed people in sales, marketing and business development, too. Prakash later said that moving from an early product to a company taught him the importance of recruiting outside his own expertise. An engineer could build the first version alone. A business required a cast.
“Even a sensationally useful product won't distribute itself.”Vipul Ved Prakash, on advice to founders in 2013
He was blunt with younger founders about distribution. They could spend too many hours adding features and too few thinking about how anyone would find the result. He also spoke about hiring for both competence and attitude, and about the almost excessive focus a young company can demand. None of this has the romance of a clever algorithm. It is a founder describing the unglamorous work between a promising demonstration and a service that survives.
The internet, indexed in real time
The next challenge involved a different kind of unwanted overflow: an endless public conversation. Prakash co-founded Topsy Labs to search and analyze social data. In a 2013 interview, he said its index contained 250 billion tweets created by a community of 200 million people. The numbers belonged to that moment, when Twitter's stream looked impossibly large and still held clues about what people cared about right now.
Topsy's job was to make the stream searchable and meaningful. The premise again depended on what people did together, although the data was no longer an anti-spam vote. It was public speech arriving by the second. Prakash talked at the time about the opportunities in connecting datasets and presenting complex information so that its patterns became visible. He was interested in both the pipes and the view out of the window.
Apple acquired Topsy in 2013. At Apple, Prakash worked on search and deep learning, a pairing that would shape his next company. He later recalled that his team built an open domain question answering system in about a week. Its performance stood out beside a much more intricate information retrieval approach. For an engineer trained on the elaborate machinery of search, the comparison was a glimpse of what learning systems might become.
Apple also gave him a product lesson. Its technology could be complicated behind the screen while the experience felt fluid in front of it. He has described that simplicity as an influence on Together AI's developer tools. Access to a powerful model is only half an invitation if using it requires a small army of infrastructure specialists.
Four founders and a larger machine
In June 2022, Prakash started Together AI with Chris Ré, Percy Liang and Ce Zhang. He had been talking with Ré about reducing the cost of building AI models. The expensive part was not merely writing code. Training and serving models demanded specialized chips, energy, networking and the knowledge to make them work efficiently. These costs could concentrate power in a small number of companies.
The way he tells the formation story has one wonderfully human detail. Ré introduced him to Zhang, whose research concerned distributed systems and moving data efficiently between machines. Within five minutes of their conversation, Prakash said, he thought they would start a company. It is a founder's compressed memory, of course, but it captures the match: a longstanding interest in open systems meeting the expertise required to run them at scale.
The founding group brought together research and industry experience. Together AI released open models, datasets and tools, including RedPajama, and built a cloud platform for training, fine-tuning and inference. The vocabulary can sound forbidding. The practical aim is easier to see: a developer should be able to choose a model, adapt it to a task and run it without first building a data center. Prakash has said he wants development to become more serverless, with less commitment to expensive computing before an experiment can begin.
2023
2025
2026
The funding milestones show how quickly the scale changed. Together announced a $20 million seed round in 2023, a $305 million Series B in 2025 and an $800 million Series C in July 2026. The latest announcement also described commitments for more than 500 megawatts of compute capacity to be financed independently by new investors. Those figures are company announcements, and they describe infrastructure ambition rather than a person's wealth. They also make the original open systems promise harder, and more consequential, to deliver.
What openness costs
Prakash is under no illusion that a model becomes broadly usable simply because its weights are available. Chips must be bought, clusters assembled, software optimized, and developers given a coherent way to use all of it. In a 2024 essay, he compared the present AI moment with the early history of personal computing. A remarkable invention needed decades of supporting systems before it could live in a pocket. He argued that AI would likewise depend on advances across hardware, software and infrastructure.
This helps explain the shift from Razor to Together. The scale has changed from messages to models, and from servers to supercomputers, but the constraint is familiar. Shared work needs a practical place to happen. In 2003 that might mean a spam signature distributed across a network. In 2026 it can mean an AI company securing enough computing power to serve its users. Together's July partnership with Y Combinator, which established a dedicated GPU cluster for startups, puts that idea in unusually concrete form.
“I think fundamentally, Together is about open and independent AI systems.”Vipul Ved Prakash, on the company's purpose
There is a useful tension in his career. The person who helped make a freely shared spam filter spent years at Apple, a company famed for tightly controlled products. He does not describe the experience as a detour. He took from it an appreciation for polish and for hiding complexity from users. Together's ambition asks him to combine that lesson with the freedom of open models. A rough toolkit can be open and still exclude almost everyone. A smooth one can make the door look inviting, although the business of keeping it open remains difficult.
His 2013 advice to founders now reads like a commentary on his own path. Build a team that knows things you do not. Think hard about distribution. Pay attention to what customers can actually use. Those principles have survived several technological fashions. They apply to the giant cluster as readily as they did to the inbox.
Prakash has said that building Together AI feels like his dream job, and that he expects to keep doing it for a long time. The statement is striking for its lack of an exit date. From a young coder looking for uninterrupted time to a CEO negotiating the economics of AI, he has kept finding new versions of the same puzzle. There is a door, there is a crowd outside it, and there is the engineering problem of making the handle work.