The inference and post-training platform powering the AI-native era
HQ: San Francisco, CA, United States | Founded: 2019 | Employees: ~250 (2026, est.) | Stage: Late Stage VC (Series F) | Website: https://www.baseten.co
Baseten was founded in 2019 in San Francisco by Tuhin Srivastava (CEO), Amir Haghighat (CTO), Philip Howes (Chief Scientist), and Pankaj Gupta, all of whom had previously worked at companies like Uber and built ML systems at scale. Their founding insight: deploying and serving machine-learning models in production was still painfully hard — and as models grew, the serving layer would become the critical bottleneck for every AI product.\n\nBaseten built a serverless-style inference platform where companies deploy, scale, monitor, and now post-train models with minimal ops burden. The company rode the LLM wave with exceptional timing: as open-weight models (Llama, Mistral, DeepSeek, Qwen) became strong enough to rival closed APIs, enterprises and AI startups turned to Baseten to serve and customize them. Revenue exploded from roughly $30M annualized in early 2025 to $200M by December 2025 and $600M by March 2026 (Sacra), with inference volume growing 40x over the same stretch.\n\nThe company raised rapidly through 2025-2026: a $75M Series C (November 2024), a $150M Series D at a $2.15B valuation (September 2025, led by Bond), a $300M Series E at a $5B valuation (early 2026, co-led with CapitalG, IVP, NVIDIA, Altimeter, and Battery Ventures), and a landmark $1.5B Series F at a $13B valuation in June 2026 led by Altimeter, Conviction, and Spark Capital with Sands Capital, Wellington, Greylock, IVP, and others. Today Baseten's customer roster reads like a who's-who of AI-native companies — Cursor, Notion, Lovable, Harvey, HubSpot, OpenEvidence, Abridge, Decagon, Parallel — and it has become the reference infrastructure for the post-training + inference era.
Baseten has become the default inference layer for the AI-native startup wave — Cursor, Notion, Lovable, Harvey, HubSpot, Decagon, and dozens more run their models on it. The company's bet is that open-weight models plus post-training (custom fine-tunes, RL environments, proprietary-data harnesses) let every company own its intelligence, and that inference is therefore the most important layer in the AI stack. That thesis has translated into extraordinary numbers: $600M annualized revenue by March 2026 (up ~1,900% YoY), 40x inference-volume growth in a year, and a $1.5B Series F at $13B — the fourth raise in 18 months. The bull case is the gravity of the AI-native customer base and a platform that embeds deeply into customers' training pipelines; the bear case is the brutal competitive arena (Together, Fireworks, vLLM-based clouds, and the hyperscalers) and the risk that inference commoditizes faster than any single layer can defend.
Browse: Home | Blog | About | All Startups A-Z | View full profile on StartupWiki