Baseten
Gives developers a hosted endpoint for a model they have already trained.
Baseten Funding History
Baseten provides infrastructure for deploying and serving machine learning models. Its work focuses on model inference and serverless GPU deployment, supporting the operation of models within AI applications.
Company Overview
Baseten Funding Timeline
Series F
$1.5B · $13B valuation
The round funds additional compute, software, and hiring as Baseten expands from serving models into helping customers post-train and optimize them.
Led by Altimeter Capital, Conviction, Spark Capital, Sands Capital, Wellington Management. Participation included Battery Ventures, Blackbird Ventures, D.E. Shaw Ventures, Greylock, IVP, Verified Capital.
Source · baseten.co ↗Series E
US$300 million · $5B valuation
Baseten framed the round around running specialized AI models in production, where inference speed, reliability and security matter.
Led by IVP, CapitalG. Participation included NVIDIA, Altimeter Capital, Battery Ventures, BOND, BoxGroup, Blackbird Ventures, Conviction, Greylock.
Source · baseten.co ↗Series D
US$150 million · $2.15B valuation
Baseten raised the round to improve the speed, reliability and cost of inference. BOND’s Jay Simons joined the board with the financing.
Led by BOND. Participation included CapitalG, Premji Invest, Conviction, IVP, Spark Capital, Greylock.
Source · baseten.co ↗Series C
US$75 million
Baseten planned to expand its inference platform through more global deployment partners and give customers more control over where their models run.
Led by IVP, Spark Capital. Participation included Greylock, Conviction, South Park Commons, Lachy Groom.
Source · baseten.co ↗Series B
US$40 million
Baseten planned to add cloud capacity, build an orchestration layer for pipelines and queues, and develop an engine to optimize workload performance.
Led by IVP, Spark Capital. Participation included Greylock, South Park Commons, Lachy Groom.
Source · baseten.co ↗Series A
US$12 million
Baseten opened its public beta alongside the round, giving machine learning teams tools to deploy models, connect them to applications, and build interfaces.
Led by Greylock. Participation included South Park Commons, Lachy Groom.
Source · baseten.co ↗
Baseten Founding Team
Tuhin Srivastava, Amir Haghighat, Philip Howes, and Pankaj Gupta worked at Gumroad before founding Baseten to help machine learning teams turn models into software people could use. Srivastava became CEO, Haghighat became CTO, and Howes became chief scientist.
Baseten Products
Baseten Company Development
Packaging a Model Became a Deployment Workflow
Truss gave teams a way to define a model’s serving logic in Python and package the dependencies and GPU settings around it. Baseten made that workflow part of a managed inference stack and also added a route for teams that need to deploy their own container images.
Baseten Competitive Landscape
Where Infrastructure Responsibility Sits
Baseten and DeepInfra overlap in hosted model inference. Baseten offers deployment infrastructure for teams bringing their own serving workloads, while DeepInfra’s API is oriented around calling models from a hosted catalog. The distinction matters when the application needs custom serving behavior.
FAQ
How does inference infrastructure differ from training infrastructure?
Inference serves an already trained model in response to requests. Training updates model parameters. The workloads have different latency, utilization and cost constraints.
Why does GPU utilization matter for model serving?
Idle accelerator capacity still costs money. Scheduling, batching and matching hardware to demand can change serving economics without changing the underlying model.
What does cold-start latency mean for serverless inference?
It is the delay while capacity and model weights are prepared for a request. Evaluate it separately from response time once a deployment is warm.
Does Baseten’s financing reveal its inference revenue?
A financing amount describes capital raised. It is not a disclosure of revenue, customer spending or gross margin.
Why are token prices insufficient for comparing inference providers?
Model quality, latency targets, context length, throughput and reliability affect the practical cost of serving an application.
How should developers test a managed model-serving platform?
Use representative traffic and measure tail latency, failures, scaling behavior and total cost under realistic concurrency.