BASETEN

Four engineers started Baseten in 2019. Two of them had already built and sold a startup called Shape.

Those two were Tuhin Srivastava and Philip Howes. Srivastava runs the company as CEO.

Reflektive bought Shape back in 2018. Amir Haghighat came from Clover Health and Pankaj Gupta came from Uber.

They kept hitting the same wall. Getting a machine learning model out of a notebook and into a live product was miserable and slow.

So they built the plumbing to do it. Years later the bet narrowed to inference.

That just means running AI models fast and cheap at massive scale.

Companies build or fine-tune an AI model. Then they need it to answer millions of user requests without falling over.

Baseten runs that part for them. You hand over a model and Baseten turns it into an API that scales up and down on its own.

You pay for the compute you actually use. No buying GPUs.

No hiring a team to keep them alive at 3am. That is basically it.

The customers are AI companies that live or die on speed, like the ones doing voice, transcription, and image generation.

The core product is the inference platform. You deploy a model and get a production API that autoscales.

Truss is their open-source tool for packaging models so they run the same everywhere. They also offer dedicated deployments for companies that want a private setup.

More recently they pushed into training and fine-tuning, so customers can build and run models in one place. The whole thing is built around one promise.

Your model, live and fast, without you managing servers.

Baseten got lucky with timing and then pressed hard. When open-source models like Llama and DeepSeek got good, companies wanted to run them without wrestling AWS directly.

Baseten was already sitting there with the tooling. It made speed the whole pitch.

Faster responses, lower cost per request, models that do not crash under load. It also spread its compute across many cloud providers instead of one.

That let it grab spare GPUs wherever they were cheap. Revenue reportedly hit $600 million on an annualized basis by early 2026.

That is up almost 1,900% in a year.

The scary part is that Baseten does not own the ground it stands on. It rents GPUs from clouds like CoreWeave and the big providers.

If GPU prices spike or supply dries up, the margins get ugly fast. It also sits in a knife fight of a market.

Together AI, Replicate, Fireworks, and the hyperscalers all want the same customers. And there is a bigger fear.

If running your own models gets easy enough, why pay a middleman at all. Baseten's answer is to be so fast and so reliable that doing it yourself feels dumb.

So far that answer is working.

Baseten raised big and often. Early backing came from Greylock, IVP, Spark Capital, Conviction, and 01A.

Mary Meeker's BOND led the $150 million Series D in 2025 at a $2.15 billion valuation. Then things went vertical.

NVIDIA put in a reported $150 million as part of a Series E in early 2026. By June 2026 the company raised $1.5 billion in a Series F at a $13 billion valuation, with Altimeter Capital among the backers.

When NVIDIA and the AI-infra crowd both want in, that tells you where the smart money thinks this is going.