Baseten’s new Baseten Training platform marks the company’s most ambitious pivot yet, moving beyond its core inference business into the increasingly contested arena of AI model training. Built on a multi‑cloud management layer, the service automatically provisions NVIDIA H100 or B200 GPUs across multiple cloud providers, enabling sub‑minute job scheduling and automated checkpointing that protects against node failures. By allowing customers to download and own their model weights, Baseten sidesteps the restrictive terms that many fine‑tuning providers impose, offering the flexibility that has become a key differentiator for enterprises eager to reduce dependence on proprietary APIs.
The platform’s technical depth is matched by its developer‑centric design. Baseten delivers per‑GPU metrics, granular checkpoint tracking, and a refreshed UI that surfaces infrastructure‑level events—all wrapped in an API that can be programmatically orchestrated by partner tools such as Oxen AI. Early adopters report dramatic savings: Oxen’s AlliumAI client achieved 84 % cost reduction, while Parsed cut end‑to‑end latency by 50 % and launched over 500 training jobs in under 48 hours. Beyond cost and speed, Baseten’s training infrastructure also fuels its inference edge; the company’s model‑performance team uses the platform to train “draft models” for speculative decoding, a technique that has delivered 650+ tokens per second on GPT‑OSS 120B.
Baseten’s strategy is clear: by owning the full AI stack—from training through inference—it can lock in customers through performance, not contracts. The move comes at a time when open‑source models are closing the gap with proprietary offerings, and enterprises are looking for reliable, cost‑effective ways to fine‑tune models for niche domains. With a $150 M Series D and a $2.15 B valuation, Baseten is positioned to invest heavily in both training and inference, aiming to become the go‑to infrastructure for companies that need custom models at scale.
Key takeaway: Baseten’s strategy of keeping customers in control of weights while providing low‑level, multi‑cloud training infrastructure aims to lock in enterprises through superior inference performance rather than restrictive contracts.
💡 Key Insight
Baseten’s strategy of keeping customers in control of weights while providing low‑level, multi‑cloud training infrastructure aims to lock in enterprises through superior inference performance rather than restrictive contracts.
Want the full story?
Read on VentureBeat →