Tuhin Srivastava and the years before the rush

At eight in the morning, late in 2019, Tuhin Srivastava arrived at Greylock with a problem he wanted to make cheaper. He had recently come out of the acquisition of his previous startup. There was time to consider another company, and he had already decided where to look: machine learning. The obstacle was the long, expensive journey between making a model and getting someone to use it.

His pitch had little patience for ceremony. Teams needed a shorter development cycle. Otherwise, he argued, most would never ship. The meeting with Sarah Guo lasted hours. A diagram from that conversation survived in her notes; over the following two years, she watched him draw versions of it again. Some founders have an elevator pitch. Srivastava apparently required a writing surface.

Seven years later, the company that grew out of those conversations sits beneath products used by people who may never learn its name. Baseten runs AI models in production. In June 2026 it announced a $1.5 billion financing at a $13 billion valuation. Those are company figures, and they describe a very different scale from the early morning meeting. The underlying question remains recognizable: how do you get useful intelligence into someone's working day?

“Otherwise, most teams are never going to ship.”

Tuhin Srivastava, on the problem behind Baseten

Fraud, feedback, and the work after the model

Srivastava's route to that question passed through Gumroad, the marketplace that lets creators sell directly to customers. He worked there as a data scientist. Philip Howes, later his co-founder, worked on machine learning too. Their responsibilities included fraud, abuse, and content moderation. Getting a prediction was one part of the job. Making it useful required them to become full-stack engineers.

Consider the fraud example Srivastava used in an early conversation about Baseten. A transaction arrives. A model flags it. Someone reviews the result. A decision can trigger a refund, a database update, or a restriction on an account. Each step needs to connect to the next. The model has delivered a judgment; the business still needs an entire little railway to carry it somewhere.

He returned to the human part of that process: decisions should feed back into the data and then into the model. A system that merely produces an answer misses the chance to learn from what happens afterward. This is a useful clue to his later interests. Even when discussing software infrastructure, Srivastava keeps circling back to the person trying to get a task done.

There was another company before Baseten. Srivastava and Howes co-founded Shape, which was acquired by Reflektive. In early 2019 they joined South Park Commons, a community where technologists work through what to build next. The setting offered room for exploration before a company had a crisp product description. It also placed them among people studying a technology whose commercial shape was still unsettled.

A company early enough to be inconvenient

At South Park Commons, founders and researchers were reading linear algebra textbooks and debating whether transformer-based models could scale. BERT and GPT-2 were objects of experimentation. Srivastava and Howes saw possibilities for customized intelligence in ordinary business workflows. It was a plausible future. Finding enough customers living in that future was another matter.

Baseten began in 2019. In its original form, it offered a broad collection of tools: model deployment, code to connect models with other services, and an application builder. Data scientists could move beyond a notebook and make something other people could interact with. This made sense given the founders' experience. They had done the connecting work themselves and knew how much of it could sit between an idea and a usable product.

In 2021, Srivastava described another consequence of reducing that work: making failure affordable. If an experiment needed weeks of engineering before anybody could try it, a team might never attempt it. Put a simple prototype in front of users sooner, and more ideas could receive an honest test. His ambition included the projects that would be abandoned. A cheap failed experiment still teaches something; an expensive unstarted one has remarkably poor manners.

The team's enthusiasm also had a playful side. Srivastava later wrote about entering too many Kaggle competitions and experimenting with OpenAI Gym, the toolkit released in 2016. The exercises included teaching simulated carts to climb mountains. These were builders attracted to the possibilities of machine learning well before inference became a popular investment category. A virtual cart does not sign a purchase order, however diligently it climbs.

Three products became one

By Srivastava's later account, Baseten spent roughly three years with a team of ten to twenty people. They were making components that would prove useful, but demand lagged behind the work. He remembered parts of that period as enjoyable: they were having a good time building. The recollection resists the tidy founder memoir in which every quiet year must be a trial endured for the benefit of the eventual ending.

Then the market changed. ChatGPT arrived in November 2022; open models such as Stable Diffusion and Whisper also brought new possibilities within developers' reach. Applications needed somewhere to run their models. Srivastava's written account places the market's arrival in 2022. His later interview describes demand gathering through 2023. The two perspectives make sense together: a change in expectations can be sudden while confidence in a business develops slowly.

Baseten made a consequential edit. It removed the application builder and the serverless code component, concentrating on deployment and inference. Years of work had produced three connected offerings. The team chose to deepen one. That is an awkward kind of progress to put in a product announcement: fewer things to sell, and considerably more responsibility for the thing that remains.

The choice changed the practical meaning of the company. Its customers were building AI applications whose responsiveness depended on the infrastructure beneath them. Baseten needed to run those models quickly and reliably as usage grew. The work became narrower in description and deeper in execution. A sentence about serving models can conceal rather a lot of engineering.

The machinery beneath an answer

In February 2025, Srivastava laid out three interlocking areas of work. First came model performance: making the software run efficiently on particular chips. Second came infrastructure that could scale across nodes, regions, and clouds. Third came model management and observability, so developers could understand what their deployments were doing. The list explains why simply obtaining GPUs leaves much of the job unfinished.

His experience at Gumroad provides a useful way to read those priorities. An answer has to arrive where it is needed, in time to be useful, with enough visibility for someone to trust the process. Now those requirements apply to larger models and more demanding applications. The engineering details have changed. The distance between a promising result and a dependable service remains his business.

The funding followed the demand. Baseten announced a $75 million Series C in February 2025 and a $150 million Series D that September. Its January 2026 Series E announcement reported $300 million raised at a $5 billion valuation. By June, the Series F announcement put the valuation at $13 billion. The numbers measure investors' commitments to a company; the product still has to earn its place one workload at a time.

Alongside that June financing, Baseten reported revenue growth of twentyfold and inference-volume growth of fortyfold over the preceding year. Its named customers included Cursor, Notion, Lovable, Harvey, HubSpot, and Decagon. These figures were the company's own account of its business. For Srivastava, the operating challenge was keeping an infrastructure supplier moving at the speed of customers whose products were themselves changing quickly.

A shorter distance to the customer

When asked how a small company moves fast, Srivastava emphasized the feedback loop with customers: hearing a problem, shipping a change, seeing what happens, and improving it. Baseten works closely with customers through forward-deployed engineering. The arrangement brings engineers into the practical details of the applications they support. A request arrives with context, rather than as an isolated item in a queue.

He also cautioned developer-tool founders about scaling before their abstractions were ready. Once people depend on an interface, changing foundational choices becomes difficult. Keep an early group small enough to learn, he advised, and use the judgment that comes from being a developer. Speed has two clocks here: the speed of finding out what is wrong, and the much slower speed of repairing a decision that thousands of users now depend on.

The early diagram and the later customer work belong in the same portrait. Srivastava repeatedly describes the steps around a model, the handoffs, and the points where learning gets delayed. His public explanations tend to return to concrete workflows. That gives his account of ambition a practical texture: remove enough friction and people can attempt something they previously left alone.

The loop keeps getting larger

In May 2026, Baseten introduced Loops, a software development kit for post-training models and moving them into production inference. It addresses a familiar break in the workflow: training on one platform and serving on another. Bringing those activities together extends the company's reach into the work that happens before a model is deployed. The name suits a business repeatedly concerned with what happens on the next pass.

September brought another extension. Baseten acquired Blaxel, whose technology supplies isolated execution environments, persistent storage, and connectivity for AI agents. A model can generate code or decide on an action. An agent needs somewhere to execute that code and retain its working state. Combining the systems is intended to bring the execution environment closer to the models it calls, reducing the network distance between successive steps.

Baseten's June announcement described customers increasingly wanting to own their intelligence: models shaped by their data, evaluations, and workflows. That aspiration is larger than a catalog of models. It imagines businesses developing systems that improve through use. Srivastava's company wants to supply the infrastructure that makes those repeated attempts feasible.

The arc leads back to the eight o'clock appointment. In 2019, Srivastava wanted teams to spend less time assembling the machinery before they could learn whether an idea worked. Baseten has since cut products, raised capital, and added capabilities around a more focused core. For all the changed scale, there is still a person waiting at the end of the pipeline. His company earns its place when that person can finally get on with the work.