Baseten

Baseten has one thesis, and it says it out loud: inference is everything. Not training, not the model itself, but the unglamorous work of running a model in production fast enough, cheaply enough, and reliably enough that a real product can depend on it. That bet has aged well. The company now says its platform handles more than a billion inference requests a day across 87 clusters on 18 cloud environments.

Founded in 2019 by Tuhin Srivastava, Amir Haghighat, Phil Howes, and Pankaj Gupta, Baseten spent its early years as a quieter MLOps tool before the generative AI wave turned inference into the bottleneck of the entire industry. It rode that wave hard. In roughly 18 months it raised four rounds and, in mid 2026, closed a $1.5B Series F at a valuation reported as high as $13B.

Baseten takes a trained model and gets it serving live traffic, then keeps it fast and up when demand spikes. You bring an open source, custom, or fine tuned model. Baseten handles the GPUs, the autoscaling, the cold starts, the runtime, and the performance tuning underneath.

The pitch to engineers is that they stop babysitting infrastructure. Package a model with the open source Truss framework, push it, and get a production API with 99.99% uptime and horizontal scale. For teams with strict data rules, the same stack runs inside their own cloud or on premises rather than only on Baseten's.

The problems Baseten chases are the ones that only show up at scale. A model that needs to spin up thousands of replicas hits container images that take minutes to pull. GPU memory constraints that off the shelf runtimes don't understand. Multi tenant isolation the industry never designed for. Its own job posts describe purpose building the container runtime and storage layers for inference, led by containerd maintainers, because patching around a decade of general purpose tooling only goes so far.

The payoff it points to is concrete: cold starts cut 2 to 3x, sub 10 second startup for trillion parameter models, and no thundering herd failures when traffic bursts. This is deep systems work, not a wrapper over someone else's cloud.

Baseten sells to the teams where latency and downtime are business risk, not an inconvenience. Think AI phone calls, real time transcription, coding assistants, and clinical tools where a stalled model is a lost customer or a missed diagnosis. Its named customers skew exactly that way: Cursor, Notion, Abridge, OpenEvidence, Clay, Gamma, Writer, Hebbia, and Mercor.

Inside those companies the buyer is technical. This is a platform for ML engineers and infrastructure teams, not a point and click tool for a non technical manager, and reviewers who wandered in expecting the latter say so plainly.

The core product is Baseten Cloud, a managed home for models with performance tuning built in. Under it sits the proprietary Inference Stack: custom model runtimes, caching, and custom kernels that Baseten claims deliver higher throughput and lower latency than the defaults.

Around that core, the lineup keeps widening. Truss packages models, Chains wires together compound AI systems, Frontier Gateway serves pre optimized versions of the latest open models, and a Training product now closes the loop from fine tuning to serving. The December 2025 acquisition of reinforcement learning startup Parsed pushed it further into post training, so a customer can train, tune, and serve in one place.

Baseten makes money the way a cloud provider does: usage based, you pay for what you run. The bill splits roughly into Model APIs, where you pay per token or per request on shared pre optimized models, and Dedicated Deployments, where you reserve GPUs and autoscale them.

The commercial hook is cost. Baseten markets model access at more than 50% below comparable OpenAI rates, and its enterprise motion leans on a forward deployed engineering team that embeds with big accounts. The catch is the flip side of usage pricing: reviewers warn that autoscaling under spiky traffic can make bills hard to predict.

Baseten competes with a crowded field of inference and serverless GPU platforms, from Modal and Replicate to Together AI, Fireworks, and Anyscale, plus the hyperscalers' own serving tools. Its argument is depth. Where many rivals sit on top of standard runtimes, Baseten owns the pipeline from the moment a model is pushed to the moment a request returns, which is what lets it fix cold starts and multi tenant isolation at the root.

The forward deployed engineering team is the other moat. It is consulting glued to infrastructure, and it makes the biggest accounts sticky.

The traction numbers are the loudest part of the story. Baseten reported roughly 20x year over year revenue growth heading into its Series F, and says the platform now clears more than a billion inference requests daily across 87 clusters on 18 clouds. Investors clearly believed it: the valuation went from about $2.15B in September 2025 to $5B in January 2026 to as high as $13B by mid 2026.

The customer side backs it up. Abridge runs clinical notes in production on Baseten, and Cursor, whose coding assistant crossed $2B in ARR, routes inference through it.

Developers who like Baseten like it for the reasons it sells itself: fast, dependable model serving, a clean path from a model to a live API, and low ops overhead, helped along by a hands on engineering team. Truss gets specific praise as a clean way to package models.

The recurring gripe is money. Usage pricing plus aggressive autoscaling can produce bills that are hard to forecast, especially for dedicated deployments under bursty traffic, and a few users report slow support responses. None of it contradicts the pitch so much as name its price.

Baseten's fundraising cadence reads like a company trying to keep up with its own demand. The June 2026 Series F brought in $1.5B across two tranches at valuations of $13B and $11B, led by Altimeter Capital, Conviction, and Spark Capital, with Sands Capital and Wellington Management as co leads. That followed a $300M Series E at $5B in January 2026 and a $150M round at roughly $2.15B in September 2025.

All told, Baseten has raised more than $2B since 2019. The backer list is deep: IVP, Greylock, CapitalG, Bond, South Park Commons, Battery Ventures, Durable Capital, and D. E. Shaw Ventures among them.

Baseten frames itself as an engineer's company, and the founding story is the tell: a team that felt the pain of shipping models firsthand and decided the infrastructure was the real product. The culture the job posts describe is high agency, first principles across the whole stack, and customer obsessed, with talent pulled from Meta, Stripe, Google, NVIDIA, and Databricks.

The work is unusually low level for a startup this young. Engineers contribute to containerd upstream, write custom communication kernels, and qualify Blackwell GPU clusters before most of the industry has touched them. Roles run out of San Francisco with an in person lean, and the company is hiring hard across engineering, infrastructure, go to market, and legal.

Baseten discloses pay on many of its roles, and the spread is wide because the ladder is: individual contributors up through managers of managers all sit in the same San Francisco market. Engineering bands run from roughly $95K at the low end to about $380K at the senior end, with sales roles carrying commission driven upside into the mid $300Ks.

Every posting states the same line about equity, so ownership is part of the standard package, not a perk reserved for early hires. All disclosed figures are US dollars and base salary; equity and commission sit on top.

The headline run is hard to miss: three up rounds and an acquisition in under a year. The $1.5B Series F in June 2026 drew wide coverage for its size and $13B valuation, and the December 2025 acquisition of reinforcement learning startup Parsed signaled a push from pure inference into owning the full model lifecycle.

The links below trace the raises, the Parsed deal, and the company's own posts on the systems work underneath the pitch.